CN109513215A

CN109513215A - A kind of object matching method, model training method and server

Info

Publication number: CN109513215A
Application number: CN201811408426.5A
Authority: CN
Inventors: 刘福乐; 谢志宁; 孙钟前
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2018-11-23
Filing date: 2018-11-23
Publication date: 2019-03-26
Anticipated expiration: 2038-11-23
Also published as: CN109513215B

Abstract

The application provides a kind of object matching method, apparatus, equipment and medium, wherein, this method comprises: server obtains object set to be matched, it include at least two objects to be matched in the object set to be matched, each object to be matched includes at least one attribute information；The prediction combination of M kind is generated according to object set to be matched, wherein is included at least two queues in every kind of prediction combination, is included at least an object to be matched in any queue；Every kind of corresponding matching probability information of prediction combination is obtained by matching prediction model；Matching prediction model includes victory or defeat prediction model, which is input with the corresponding at least one attribute information of each object at least two queues, to fight victory or defeat probability as output；Corresponding matching probability information is combined according to every kind of prediction, determines object matching result.Point side for player's capabilities gap great disparity occur can be effectively prevented from using this method to generate, the matching result of optimization object battle grouping.

Description

A kind of object matching method, model training method and server

Technical field

This application involves field of network game technology more particularly to a kind of object matching methods, model training method, service Device and computer readable storage medium.

Background technique

Recently as the rapid development of e-sports industry, player also increasingly closes the fairness for class game of racing Note, wherein online tactics sports class (Multiplayer Online Battle Arena, the abbreviation MOBA) game of more people is by complete The favor of ball player.With traditional RTS (real-time strategy, real-time policy) game difference, player is needed to grasp by MOBA The role of work is reduced to the single arm of the services from groups of building, multiple arm of the services etc. --- and thus hero derives with camp hero arena Battle is the hot topic MOBA class game such as Dota, heroic alliance, heroic three states, the king's honor of core playing method.

The basic principle of sports class online game be player matches that will be on close level together, avoid the occurrence of both sides as far as possible What strength had big difference rolls situation.The matching system of the competing game of electricity at present all realizes player matches unit with player's integral algorithm 's.And ELO integral algorithm is generally used in player's integral algorithm, it is approximate that basic thought assumes that all game levels are obeyed It is distributed in the logic of normal distribution, assumes that each game level variance is fixed, game level is determined by mean value, each player's Level is by ELO points of characterizations, and in specific matching, the ELO based on team member in Liang Ge troop divides summation equal or similar principle, The Liang Zhi troop of battle is matched from match-pool, so that the player in this Liang Zhi troop participates in battle game.

Since the ELO matching system based on ELO algorithm is merely with ELO points, the actual implementation of player is portrayed from single dimension Power is caused player's integral not to be consistent with player's strength, is matched based on ELO score value so its Algorithm Error is larger with ELO algorithm The true strength of Liang Zhi troop have big difference.For example, a kind of matching result that usually will appear is, ELO points of Liang Zhi troop Summation is close, but actually this Liang Zhi troop wide margin, and a troop is all the trumpet of the stronger player of strength, and another Branch troop is all the player for reaching the score value for the first time, causes to be that small size troop rolls opponent, i.e., typically rolls office.

Therefore, the matching system based on ELO integral algorithm is based only upon ELO points and positions to game level, but it is positioned Not enough precisely, both sides' strength has big difference after matching, influences the balance of game, reduces player gaming experience.

Summary of the invention

A kind of object matching method, the method for model training and server provided by the embodiments of the present application, using game pair As possessed attribute information carrys out the battle strength of forecasting game object, so that the accuracy of game object strength prediction is promoted, And then it is advantageously implemented the reasonability of team matching.

In view of this, the application first aspect provides a kind of object matching method, this method comprises:

Obtain object set to be matched, include at least two objects to be matched in the object set to be matched, it is described to Matching object includes at least one attribute information；

The prediction combination of M kind is generated according to the object set to be matched, M is positive integer, wherein wrap in every kind of prediction combination At least two queues are included, include at least an object to be matched in any queue；

Every kind of corresponding matching probability information of prediction combination is obtained by matching prediction model；Wherein, the matching Prediction model includes victory or defeat prediction model, and the victory or defeat prediction model is corresponding extremely with each object at least two queues A kind of few attribute information is input, to fight victory or defeat probability as output；

Corresponding matching probability information is combined according to every kind of prediction, determines object matching result.

The application second aspect provides a kind of model training method, this method comprises:

Determine that the first training sample set, each sample that first training sample is concentrated include every at least two troops The corresponding at least one attribute information of a object and battle result；

Each sample training neural network is concentrated according to first training sample, victory or defeat prediction model is obtained with training, The victory or defeat prediction model is input with the corresponding at least one attribute information of each object at least two troops, with Fighting victory or defeat probability is output；

Matching prediction model is determined according to the victory or defeat prediction model.

The application third aspect provides a kind of object matching device, which includes:

Obtain module, include for obtaining object set to be matched, in the object set to be matched at least two to With object, the object to be matched includes at least one attribute information；

Prediction combination determining module, for generating the prediction combination of M kind according to the object set to be matched, M is positive integer, Wherein, include at least two queues in every kind of prediction combination, include at least an object to be matched in any queue；

Prediction module, for obtaining every kind of corresponding matching probability information of prediction combination by matching prediction model； Wherein, the matching prediction model includes victory or defeat prediction model, and the victory or defeat prediction model is with each right at least two queues As corresponding at least one attribute information is input, to fight victory or defeat probability as output；

Determining module determines object matching result for combining corresponding matching probability information according to every kind of prediction.

The application fourth aspect provides a kind of model training apparatus, which includes:

Training sample determining module, for determining the first training sample set, each sample that first training sample is concentrated This includes the corresponding at least one attribute information of each object and battle result at least two troops；

Training module is obtained for concentrating each sample training neural network according to first training sample with training Victory or defeat prediction model, the victory or defeat prediction model is with the corresponding at least one attribute of each object at least two troops Information is input, to fight victory or defeat probability as output；

Model determining module, for determining matching prediction model according to the victory or defeat prediction model.

The 5th aspect of the application provides a kind of server, which includes processor and memory:

Said program code is transferred to the processor for storing program code by the memory；

The processor is used for the object matching provided according to the above-mentioned first aspect of instruction execution in said program code Method.

The 6th aspect of the application provides a kind of server, which includes processor and memory:

The processor is used for the model training provided according to the above-mentioned second aspect of instruction execution in said program code Method.

The 7th aspect of the application provides a kind of computer readable storage medium, including instruction, when it is transported on computers When row, so that computer executes the object matching method provided such as above-mentioned first aspect, or execute as above-mentioned second aspect mentions The model training method of confession.

As can be seen from the above technical solutions, the embodiment of the present application has the advantage that

A kind of object matching method is provided in the embodiment of the present application, in the method, server first obtains to be matched right It include at least two objects to be matched in the object set to be matched as set, each object to be matched includes at least one Attribute information；The prediction combination of M kind is generated according to the object set to be matched, M is positive integer, wherein in every kind of prediction combination Including at least two queues, an object to be matched is included at least in any queue；Every kind is obtained in advance by matching prediction model It surveys and combines corresponding matching probability information；Wherein, the matching prediction model includes victory or defeat prediction model, at least two Respectively corresponding at least one attribute information is input for each object in a queue, to fight victory or defeat probability as output；Root Corresponding matching probability information is combined according to every kind of prediction, determines object matching result.In the method, it is predicted based on victory or defeat Model carries out battle victory or defeat probabilistic forecasting, and the victory or defeat prediction model is, victory or defeat prediction resulting by sample data training Mapping relations of the model learning into battle troop between the attribute information of each object and battle victory or defeat probability；Therefore, pass through The victory or defeat prediction model can directly predict the battle victory or defeat probability between above-mentioned prediction combination squadron 5 based on the mapping relations, Based on the battle victory or defeat determine the probability object matching as a result, the application can be avoided point side for player's capabilities gap great disparity occur produces It is raw, so as to the object matching result of optimization object battle grouping.

Detailed description of the invention

Fig. 1 is a kind of application scenarios schematic diagram of object matching method in the embodiment of the present application；

Fig. 2 is a kind of flow diagram of object matching method in the embodiment of the present application；

Fig. 3 is a kind of structural schematic diagram of victory or defeat prediction model in the embodiment of the present application；

Fig. 4 is a kind of structural schematic diagram of role's prediction model in the embodiment of the present application；

Fig. 5 is a kind of flow diagram of model training method in the embodiment of the present application；

Fig. 6 is the flow diagram of another model training method in the embodiment of the present application；

Fig. 7 is a kind of overall work configuration diagram of object matching method in the embodiment of the present application；

Fig. 8 is a kind of structural schematic diagram of object matching device in the embodiment of the present application；

Fig. 9 is a kind of structural schematic diagram of model training apparatus in the embodiment of the present application；

Figure 10 is a kind of structural schematic diagram of server in the embodiment of the present application.

Specific embodiment

In order to make those skilled in the art more fully understand application scheme, below in conjunction in the embodiment of the present application Attached drawing, the technical scheme in the embodiment of the application is clearly and completely described, it is clear that described embodiment is only this Apply for a part of the embodiment, instead of all the embodiments.Based on the embodiment in the application, those of ordinary skill in the art exist Every other embodiment obtained under the premise of creative work is not made, shall fall in the protection scope of this application.

The description and claims of this application and term " first ", " second ", " third ", " in above-mentioned attached drawing The (if present)s such as four " are to be used to distinguish similar objects, without being used to describe a particular order or precedence order.It should manage The data that solution uses in this way are interchangeable under appropriate circumstances, so as to embodiments herein described herein can in addition to Here the sequence other than those of diagram or description is implemented.In addition, term " includes " and " having " and their any deformation, Be intended to cover it is non-exclusive include, for example, containing the process, method of a series of steps or units, system, product or setting It is standby those of to be not necessarily limited to be clearly listed step or unit, but may include be not clearly listed or for these mistakes The intrinsic other step or units of journey, method, product or equipment.

In the prior art, matching system is generally grouped the player for participating in game fighting based on ELO points, however very In more situations, ELO points can not objectively and accurately reflect that the real gaming of player is horizontal, therefore, be based only upon the ELO points to ginseng Matched packet is carried out with each player of game fighting, it is easy to there is the situation of each group capabilities gap great disparity matched, by This will affect the balance of game, reduce the game experiencing of player.

In order to solve above-mentioned the technical problems existing in the prior art, the embodiment of the present application provides a kind of object matching side Method, this method model player's strength according to the corresponding at least one attribute information of game object using victory or defeat prediction model To predict the battle victory or defeat probability of at least two troops, and based on fighting victory or defeat probability selecting object matching as a result, it is possible to Guarantee the balance of object grouping, point side for effectively avoiding the occurrence of player's capabilities gap great disparity generates, and guarantees subsequent team matching Reasonability.

It should be understood that object matching method provided by the embodiments of the present application can be applied to service having data processing function Device, the server are specifically as follows application server, or Web server, in practical application deployment, which can Think separate server, or cluster server, the server can determine the matching result of more game fightings simultaneously.

Technical solution provided by the embodiments of the present application in order to facilitate understanding, below with reference to practical application scene to the application reality The object matching method for applying example offer is introduced.

Referring to Fig. 1, Fig. 1 is the application scenarios schematic diagram of object matching method provided by the embodiments of the present application.The applied field It include multiple terminal devices 101, game application server 102 and match server 103 in scape.Wherein, terminal device 101 is fortune The equipment that row has game application, the game application are specifically as follows APP (Application) game application, or webpage Version game application, does not do any restriction to the operation form of game application herein；The terminal device be specifically as follows smart phone, Computer, personal digital assistant (Personal Digital Assitant, PDA), tablet computer etc..

Wherein, 101 object game player of terminal device, in response to game player operation to game application server 102 send battle request, and request game application server 102 is other players that itself matching participates in same field game fighting, And determine that the battle of this game fighting divides side result；Game application server 102 is used to be sent according to each terminal device 101 Battle request, determine the player for participating in same field game fighting, and by the corresponding at least one attribute letter of these players Breath is sent to match server 103 as matching request；Match server 103 receives the transmission of game application server 102 After matching request, object matching method provided by the embodiments of the present application is executed, object set to be matched is obtained from the matching request It closes, and then determines the object matching for corresponding to the object set to be matched as a result, i.e. determining this game fighting divides side result.

When match server 103 specifically determines object matching result, is first combined according to acquired object to be matched and generate M Kind of prediction combination, wherein every kind of prediction combination includes at least two queues, to be matched right including at least one in any queue As；Then the corresponding matching probability information of every kind of prediction combination is obtained by matching prediction model, specifically, match server 103 is pre- by the victory or defeat for predicting that each corresponding at least one attribute information of object is input in matching prediction model in combination Model is surveyed, which correspondingly exports this kind prediction and combine corresponding battle victory or defeat probability, and the battle victory or defeat is general Rate combines corresponding matching probability information as this kind prediction；Finally, corresponding matching probability is combined according to various predictions Information determines object matching as a result, determining point side result of game fighting.

After match server 103 determines object matching result, which is passed through into game application server 102 be sent to participate in this game fighting each terminal device 101 so that each game player know battle point side as a result, And side result is divided to participate in game fighting according to the battle.

It should be understood that in practical applications, also can use a server execute above-mentioned game application server 102 and With operation performed by server 103, i.e., the function that above-mentioned game application server 102 and match server 103 are realized can It is realized in a server with integrated.

It should be noted that above-mentioned match server 103 utilizes victory or defeat prediction model, according to the corresponding at least one of player Attribute information models the strength of each player, and the real gaming of player can be characterized by guaranteeing the player's strength evaluated more It is horizontal；In turn, based on the player's strength so evaluated, predict that corresponding battle is combined in various predictions using victory or defeat prediction model Victory or defeat probability, with reference to the battle victory or defeat determine the probability object matching as a result, thus avoiding the occurrence of point of player's capabilities gap great disparity Side, optimization player fight the matching result of grouping.

It should be noted that object matching method provided by the embodiments of the present application is suitable for the various sports grounds for needing to be grouped Scape, specifically, the object matching method can be applied to a plurality of types of game such as MOBA class, Grapple, tackling key problem class, strategic class In scene, any restriction is not done to the scene of game of object matching method application herein.Above-mentioned application scenarios shown in FIG. 1 are only For a kind of example, in practical applications, object matching method provided by the embodiments of the present application can also be applied to other application field Scape does not do any restriction to the application scenarios of the object matching herein.

Object matching method provided by the present application is introduced below by embodiment.

Referring to fig. 2, Fig. 2 is a kind of flow diagram of object matching method provided by the embodiments of the present application.For the ease of Description, following embodiments are described scheme using server as executing subject.As shown in Fig. 2, the object matching method packet Include following steps:

Step 201: object set to be matched is obtained, includes at least two objects to be matched in the object set to be matched, Each object to be matched includes at least one attribute information.

In the application scenarios of MOBA game, object to be matched refers to participating in the player of game fighting, and server is from hair The each player for choosing in the player of battle request and participating in same field game fighting is played, these players constitute this game fighting Corresponding object set to be matched；The each player corresponding at least one for participating in this game fighting is obtained from database Attribute information.It should be understood that the corresponding at least one attribute information of player herein is that can reflect player to a certain extent Play level information.

It should be understood that the player for participating in a game fighting usually has V*N, wherein V indicates the group's number divided, that is, swims Play battle divides number of edges mesh, and N indicates player's number in each group；It correspondingly, include V*N object for appreciation in object set to be matched Family, this V*N player respectively correspond at least one attribute information.

Optionally, above-mentioned at least one attribute information can specifically include: the whole competitive state information of object, object Recent competitive state information, the ELO ranking system point of object, the game section of object, object personal information at least one Kind；Wherein, the personal information of object includes the natural quality information of object and/or the game asset information of object.

The whole competitive state information of object generally includes all game fighting plays and all trips that the player participates in The corresponding victory or defeat probability of play battle play.The recent competitive state information of object generally includes the game pair that the player participates in the recent period Battlefield time and the corresponding victory or defeat probability of these game fightings participated in the recent period, specifically, can be by player in nearly one month The game fighting play of participation and the corresponding victory or defeat probability of the game fighting participated in this month are believed as recent competitive state Breath, can also the game fighting play of participation and the corresponding victory or defeat of game fighting participated in this week in nearly one week by player Probability is as recent competitive state information, it is of course also possible to which the period of random length is arranged according to the actual situation as recent The competitive state information corresponding period.The ELO ranking system of object point is ELO points, is each field participated according to player The score value that the victory or defeat result of game fighting is determined.The game section of object refers to the game ratings of player, can be according to object for appreciation Victory or defeat result of the family in each field game fighting and correspondingly change.The personal information of object specifically includes the natural quality of object The game asset information of information and/or object, the natural quality information of object refer to player's upload when registering game account Basic personal information, such as gender, age, the game asset information of object refers to the game money that player has in gaming It produces, such as game gold coin, game article.

It should be noted that when assessing the play level of player, the whole competitive state information of object and object it is close Phase competitive state information reference value with higher, the i.e. whole competitive state information of references object and the recent sports of object The play level that status information evaluates more is consistent with the true strength of player.Therefore, the corresponding category of object that server obtains Property information typically at least includes the whole competitive state information of object and the recent competitive state information of object, to utilize object Whole competitive state information and recent competitive state information more truely and accurately assess the play level of player.

In one possible implementation, server can be obtained from the matching request that game application server is sent Object set to be matched.Specifically, server receives the matching request that game application server is sent, include in the matching request The corresponding at least one attribute information (N is positive integer) of 2N objects to be matched, the 2N objects to be matched are according to right The game section and/or ELO of elephant point determination；In turn, server obtains object set to be matched from matching request, should to With including 2N objects to be matched in object set.

It should be noted that this kind of implementation is primarily directed to the scene of game of two sides battle, therefore, above-mentioned matching It include the corresponding at least one attribute information of 2N object in request, N indicates a player for including in each party fought Number.It should be understood that should include 3N object corresponding at least one in above-mentioned matching request for the scene of game of tripartite's battle Attribute information should include that 4N object is corresponding at least for the scene of game of four directions battle, in above-mentioned matching request A kind of attribute information, and so on.

In practical applications, game application server usually can receive a large amount of players within the same period and pass through end The battle request that end equipment is initiated, these fight the account identification that player is usually carried in request；Game application server connects After the battle request for receiving player's initiation, need to be that each player matches for initiating battle request are participated according to game fighting rule Other players of same field game fighting choose the same field of participation from the player for initiating battle request during this period of time and swim Play the player fought, for example, it is assumed that game rule is that player is divided into two teams in a game fighting, every team includes N number of player, Then game application server correspondingly chooses 2N player from the player that this period initiates battle request and participates in same field Game fighting；When game application services the player of the same field game fighting of implement body selection participation, it will usually according to each battle The account identification carried in request determines the corresponding game section of each player and/or ELO points, in turn, chooses game section phase It is together or close, and/or, ELO split-phase is together or similar 2N player is as the player for participating in same field game fighting.

After game application server selects the player for participating in same field game fighting, accordingly based upon participation this game The account identification of each player of battle transfers the corresponding at least one attribute information of these players from database, and The corresponding at least one attribute information of these players is added to be sent in matching request and is used to determine matching result Server；After determining that the server of matching result receives the matching request that game application server is sent, according to this Object set to be matched is determined with request, it is corresponding at least further according to each object to be matched in the object set to be matched A kind of attribute information, determines the matching result of game fighting, that is, determine game fighting divides side result.

It should be understood that in practical applications, the function and be used to determine game fighting that above-mentioned game application server is realized The function that server with result is realized, which is desirably integrated on same server, to be realized, it can You Yitai server determines ginseng With all players of same field game fighting, and determine that the player for participating in this game fighting divides side result.

It should be understood that the mode of above-mentioned acquisition object set to be matched is only a kind of example, in practical applications, server Object set to be matched can be obtained by other means, object to be matched is not obtained to server herein and combined taken side Formula does any restriction.

Step 202: M kind prediction combination (M is positive integer) being generated according to object set to be matched, wherein every kind of prediction group Include at least two queues in conjunction, includes at least an object to be matched in any queue.

After server gets object set to be matched, generated according to game fighting rule and the object set to be matched The prediction combination of M kind, wherein it include at least two queues in every kind of prediction combination, it is to be matched including at least one in any queue Object, queue herein can be understood as grouping or point side in game fighting.

It should be understood that the object number for including in the number of queues and each queue that include in above-mentioned every kind of prediction combination is equal It is related to game fighting rule, wherein ginseng specified in the number of queues and game fighting rule for including in every kind of prediction combination It is equal with group's number of battle, i.e., divide number of edges equal to game fighting, the object number and game rule for including in each queue Specified in include in each group number of players it is equal.For example, it is assumed that game fighting rule is 5V5 confrontation (i.e. by respectively Liang Zhi troop including 5 players fights), then in every kind of prediction combination that server is generated according to object set to be matched Include two queues, includes 5 players in each queue.

It is pre- determining in order to which the prediction combination for guaranteeing that server is determined can cover all combining forms being likely to occur When surveying combination, server can determine prediction combination using number of combinations calculation formula.Specifically, server first determine it is to be matched right As the object total number of object to be matched in set passes through group then according to membership in object total number and default team It closes number calculation formula and determines the prediction combination of M kind, which is number of combinations.

It should be understood that the object total number for including in above-mentioned object set to be matched actually be set to be matched in include Player's total number；Membership is determined according to game fighting rule in default team, every specified in game fighting rule The number of players for including in a group is membership in the default team.

Server according in object set to be matched object total number and default team in membership, using such as formula (1) number of combinations calculation formula shown in calculates total kind of several M of prediction combination.

M=C (V*N, N) * C ((V-1) * N, N) * C ((V-2) * N, N) * ... * C (1*N, N) (1)

Wherein, M indicates total kind of number of prediction combination；The form of calculation of C () expression number of combinations；V is indicated in game fighting Number of queues, i.e. game fighting divide number of edges；N indicates the object number that each queue is included.It should be understood that the M so determined Kind prediction combination is different, i.e., the object that each queue in every kind prediction combination is included is all different.

For example, server is directed to 5V5 dual meet, C (2*5,5) * C (5,5)=252 kind can be calculated using formula (1) Prediction is combined, and is included 5 queues in every kind of prediction combination, is included 5 objects in each queue；In another example being fought for 3V3V3 Match (3 troops by respectively including 3 players fight), using formula (1) can calculate C (3*3,3) * C (2*3, 3) * C (3,3)=1680 kinds of prediction combination.

It should be understood that above-mentioned determination predicts that combined mode is only a kind of possible implementation, and in practical applications, service Device can also determine prediction combination according to object set to be matched using other modes, for example, server can be according to game pair War rule is randomly grouped each player in object set to be matched and determines several prediction combination, herein not Predict that combined mode does any restriction to determining.

Step 203: obtaining the corresponding probability match information of every kind of prediction combination by matching prediction model；Wherein, this It include victory or defeat prediction model with prediction model, which is corresponding with object each at least two queues At least one attribute information is input, to fight victory or defeat probability as output.

After server generates the prediction combination of M kind according to object set to be matched, using matching prediction model determine every kind it is pre- It surveys and combines corresponding matching probability information, which can characterize its corresponding prediction combination and carry out game fighting Battle is as a result, such as victory or defeat result.When concrete application, server respectively corresponds to each object to be matched in every kind of prediction combination At least one attribute information be input in the matching prediction model, matching prediction model accordingly based upon prediction combination in it is each The corresponding at least one attribute information of object to be matched determines that corresponding to this kind predicts combined matching probability information.

It should be noted that including victory or defeat prediction model in above-mentioned matching prediction model, the victory or defeat prediction model is at least The corresponding at least one attribute information of each object is input in two queues, to fight victory or defeat probability as output.Specifically In application, the corresponding at least one attribute information of object to be matched each in every kind of prediction combination is input to by server After in prediction model, victory or defeat prediction model therein will correspondingly obtain this kind and predict that each object to be matched is respectively in combination It is corresponding at least one attribute information, in turn, the victory or defeat prediction model according to the prediction combine in each object to be matched respectively Corresponding at least one attribute information determines that corresponding victory or defeat probability is combined in this kind prediction.

It should be understood that the corresponding at least one attribute information input victory or defeat of object to be matched each in prediction combination is pre- When surveying model, need clearly to distinguish which queue the corresponding at least one attribute information of each object to be matched specifically corresponds to, So that victory or defeat prediction model can be according to the corresponding at least one attribute of each object to be matched in each queue Information predicts the corresponding battle strength of each queue in prediction combination, and then predicts that each queue carries out in prediction combination The victory or defeat probability of battle.

Specifically, in order to clearly distinguish the specific corresponding team of the corresponding a variety of attribute informations of each object to be matched Column, server can be before by each corresponding attribute information input victory or defeat prediction models of object to be matched, first will be every A corresponding at least one attribute information of object to be matched is correspondingly divided in the queue belonging to it, i.e., will belong to same The corresponding at least one attribute information of each object to be matched of a queue takes together, as the corresponding attribute of the queue Information；In turn, the corresponding attribute information of each queue is inputted into victory or defeat prediction model.For example, it is assumed that including team in prediction combination Column 1 and queue 2, include object a, object b and object c in queue 1, include object d, object e and object f in queue 2, will be pre- It surveys before the corresponding at least one attribute information input victory or defeat prediction model of each object in combination, server first will be right As a, object b and the corresponding attribute information of object c take together as the corresponding attribute information of queue 1, by object d, right As e and the corresponding attribute information of object f take together as the corresponding attribute information of queue 2, in turn, server is by queue 1 corresponding attribute information and the corresponding attribute information of queue 2 input victory or defeat prediction model, so that the victory or defeat prediction model is according to team The corresponding attribute information of column 1 and the corresponding attribute information of queue 2 predict that corresponding victory or defeat probability is combined in the prediction.

It should be understood that it is above-mentioned summarize attribute information as unit of queue in the way of be only a kind of possible implementation, in reality In the application of border, server can also distinguish the corresponding at least one category of each object to be matched using modes such as addition queue identities Property the specific corresponding queue of information, any restriction is not done to the mode for distinguishing the affiliated queue of attribute information herein.

It should be noted that the battle victory or defeat probability of above-mentioned victory or defeat prediction model prediction output may include in prediction combination The probability of this battle is won in each queue；For example, it is assumed that a certain prediction combination includes queue 1, queue 2 and queue 3, victory or defeat is pre- Model prediction is surveyed in the battle of queue 1, queue 2 and queue 3, the triumph probability of queue 1 is 21%, the corresponding triumph of queue 2 Probability is 50%, and the corresponding triumph probability of queue 3 is 29%, then 21%, 50% and 29% is the output of victory or defeat prediction model Fight victory or defeat probability.In the scene of game of two sides battle, the battle victory or defeat probability of above-mentioned victory or defeat prediction model prediction output can To be probability that any side wins battle；For example, it is assumed that a certain prediction combination includes queue 1 and queue 2, mould is predicted in victory or defeat Type predicts that in the battle of queue 1 and queue 2, the triumph probability of queue 1 is 60%, then the output of victory or defeat prediction model to defeating Negative probability is 60%, and correspondingly, triumph probability of the queue 2 in this fights is 40%.

The working principle of above-mentioned victory or defeat prediction model in order to facilitate understanding, below by taking the scene of game of two sides battle as an example, The working principle of the victory or defeat prediction model is introduced in conjunction with the model structure of victory or defeat prediction model:

Referring to Fig. 3, Fig. 3 is the structural schematic diagram of victory or defeat prediction model provided by the embodiments of the present application.As shown in figure 3, will In prediction combination after the corresponding at least one attribute information input victory or defeat prediction model of each player of each queue, victory or defeat Deep-neural-network (Deep Nerural Network, DNN) hidden layer in prediction model corresponding to each player is according to each The corresponding at least one attribute information of player, extracts the player characteristic that can correspondingly reflect player level；Then, By the corresponding player characteristic of each player be input to its belonging to the corresponding DNN hidden layer of troop, DNN corresponding to each troop Hidden layer is correspondingly determined to troop's feature of reflection troop's strength according to the player characteristic of input；In turn, by each troop pair The troop's feature answered is input to match DNN hidden layer 1 and match DNN hidden layer 2, right using match DNN hidden layer 1 and match DNN hidden layer 2 The strength of each troop behind point side is indicated respectively and depth modelling, finally by Softmax layers of output victory or defeat probability.

It should be understood that the structure of above-mentioned victory or defeat prediction model shown in Fig. 3 is only a kind of example, and in practical applications, the victory Negative prediction model can also be the neural network model of other structures, not do any limit to the structure of victory or defeat prediction model herein It is fixed.

In many MOBA scene of game, the player for participating in game is also an option that oneself hero angle used in battle Color, different heroic roles have different game skills.However, many MOBA game rules define in the same troop Do not allow that there are identical heroic roles.Therefore, in a game fighting, player can not probably choose oneself and want The heroic role used, or even many players can only select the heroic role being bad at using itself in battle, so will Substantially reduce the game experiencing of player.

It, can also be further in matching prediction model provided by the embodiments of the present application in order to avoid above situation as far as possible In add role's prediction model, using role's prediction model predict each player in battle may selection role, in turn When determining object matching result, the player that will likely select same role is avoided to be divided to the same troop, to guarantee to play Family can participate in battle using the heroic role for oneself wanting to play, and improve the game experiencing of player.

It should be noted that role's prediction model provided by the present application is used in one kind is fought with object at first T innings Role sequence is classified as input, fights used role with T+1 innings of object participation and role's probability of occurrence is the nerve of output Network.When concrete application, server, can before the role that may be selected in battle using role's prediction model prediction player With elder generation from obtaining player role's sequence for selecting in first T innings battle in offline database, and then by role's sequence inputting Role's prediction model, with the role that may be selected using role's prediction model prediction player at T+1 innings.

It should be understood that above-mentioned T+1 innings of battles refer specifically to the current battle to be participated in of player, correspondingly, first T innings refers to Field the T battle that be player participate in front of participating in this battle.The T value can determine according to actual needs, can usually set It is set to 10 or 20, naturally it is also possible to be set as other values according to actual needs, not be specifically limited herein to T value.

It should be noted that role's sequence that above-mentioned role's prediction model is used according to player at first T innings, may only export One role, it is also possible to while exporting multiple roles；If role's prediction model only exports a role, illustrate player in T+ It is 1 innings very big a possibility that using the role；If role's prediction model exports multiple roles simultaneously, illustrate player at T+1 innings A role may be chosen from this multiple role and participates in local exchange battle, and server can be further defeated according to role's prediction model The corresponding probability of occurrence of these roles out determines that player selects the probability of each role.

The working principle of above-mentioned role's prediction model in order to facilitate understanding, below with reference to the model structure of role's prediction model The working principle of role's prediction model is introduced:

Referring to fig. 4, Fig. 4 is the structural schematic diagram of role's prediction model provided by the embodiments of the present application.As shown in figure 4, will Player's role's sequence inputting role's prediction model used in first T innings battle, i.e., the role used player at the t1 moment, The role that uses at the t2 moment ..., in the role that the tT moment uses input role's prediction model respectively, predicted through the role Shot and long term memory network (Long Short-Term Memory, LSTM) hidden layer of model, DNN hidden layer 1, DNN hidden layer 2, DNN are hidden Role's sequence for using at first T innings of 3 couples of player of layer is handled, most afterwards can at T+1 innings through the softmax layers of output player The role that can be used and role's probability of occurrence.

It should be understood that the structure of above-mentioned role's prediction model shown in Fig. 4 is only a kind of example, the role in practical applications Prediction model can also be the neural network model of other structures, not do any restriction to the structure of role's prediction model herein.

When matching in prediction model includes victory or defeat prediction model and role's prediction model, server can will utilize victory or defeat Prediction model is according to the victory or defeat that each corresponding at least one attribute information of object to be matched is predicted in prediction combination Probability, as the first matching probability information；By the role used in being fought according to object at first T innings using role's prediction model Role's prediction result that sequence prediction obtains, i.e. object role and role's probability of occurrence used in T+1 innings are as Two match informations.

It should be understood that in practical applications, server can determine the first matching probability information first with victory or defeat prediction model, The second matching probability information is determined using role's prediction model afterwards；Server can also determine second first with role's prediction model Matching probability information determines the second matching probability information using victory or defeat prediction model afterwards；Certainly, server can also utilize simultaneously Victory or defeat prediction model determines the first matching probability information, determines the second matching probability information using role's prediction model.Herein not Any restriction is done to the job order of victory or defeat prediction model and role's prediction model.

It should be noted that in practical applications, victory or defeat prediction model and role's prediction model can integrate as one in advance Model is surveyed, the prediction model is to predict the corresponding at least one attribute information of each player and each player in combination As input, combine corresponding victory or defeat probability and each player with prediction makes the role's sequence used at first T innings at T+1 innings The role that uses and role's probability of occurrence are as output.

Step 204: corresponding matching probability information being combined according to every kind of prediction, determines object matching result.

It is respective according to every kind of prediction combination after server gets every kind of corresponding matching probability information of prediction combination Corresponding matching probability information, judges whether each queue in various prediction combinations meets battle condition, that is, judges various pre- In the reasonable scope whether the capabilities gap between each queue in survey combination, and then combine from prediction according to the judgment result Middle selected target matching result thereby determines that game fighting point side result.

When matching in prediction model only includes victory or defeat prediction model, matching probability information is correspondingly predicted for the victory or defeat The battle victory or defeat probability of model output, then server is when determining object matching result, according to the every of victory or defeat prediction model output Corresponding battle victory or defeat probability is combined in kind prediction, judges that every kind of prediction combines whether corresponding battle victory or defeat probability meets default victory Negative probability screening conditions meet the default victory or defeat probability screening conditions and then illustrate strength in prediction combination between each queue At gap in the reasonable scope, each queue in prediction combination carries out a possibility that rolling situation occur to wartime lower.

It should be understood that above-mentioned default victory or defeat probability screening conditions can be set as at the victory or defeat probability of victory or defeat prediction model output In in a certain preset range；For example, default victory or defeat probability screening conditions can be set as the scene of game of two sides battle The battle victory or defeat probability of victory or defeat prediction model output is between 48% to 52% if corresponding victory or defeat probability is combined in certain prediction Between 48% to 52%, then illustrates that the strength of prediction combination Zhong Liangge troop is substantially suitable, correspondingly determine that the prediction is combined Meet default victory or defeat probability screening conditions；In another example default victory or defeat probability can be screened for the scene of game of tripartite's battle Condition is set as the corresponding battle victory or defeat probability of each troop of victory or defeat prediction model output between 30% to 35%, if in advance It surveys the corresponding battle victory or defeat probability of three troops in combination to be between 30% to 35%, then illustrates in prediction combination three The strength of troop is substantially suitable, correspondingly determines that prediction combination meets the default victory or defeat probability screening conditions.

It should be understood that above-mentioned default victory or defeat probability screening conditions can be configured according to actual needs, for different trips Play battle scene, can correspondingly be arranged different default victory or defeat probability screening conditions, not screen herein to default victory or defeat probability Condition does any specific restriction.

When with specific reference to victory or defeat determine the probability object matching result, server can first judge that various prediction combinations are respectively right Whether the battle victory or defeat probability answered meets default victory or defeat probability screening conditions, and it is general that battle victory or defeat is filtered out from all predictions combination Rate meets the prediction combination of the default victory or defeat probability screening conditions, if meeting the prediction combination of the default victory or defeat probability screening conditions There are multiple, then server can also further select the prediction combination that most can guarantee battle balance from these prediction combinations As object matching result.

For example, in the scene of game of two sides battle, it is assumed that the battle victory or defeat probability characterization of victory or defeat prediction model output is pre- The probability that one party in combination obtains battle triumph is surveyed, prediction 1 corresponding battle victory or defeat probability of combination is 43%, prediction combination 2 Victory or defeat probability be 80%, prediction combination 3 corresponding battle victory or defeat probability be 52%, preset victory or defeat probability screening conditions be victory or defeat The battle victory or defeat probability of prediction model output is between 40% to 60%；Prediction combination 1 and prediction group can be determined by judgement It closes 3 and is all satisfied default victory or defeat probability screening conditions, it is in turn, corresponding according to prediction 1 corresponding victory or defeat probability of combination and prediction combination 3 Battle victory or defeat probability, determine that the prediction 3 battle balances fought of combination are more preferable, therefore, can will prediction 3 conducts of combination Object matching result.

When matching in prediction model includes simultaneously victory or defeat prediction model and role's prediction model, matching probability information is specific Including the first matching probability information (the battle victory or defeat probability of victory or defeat prediction model output) and the second matching probability information, (role is pre- Survey role's prediction result of model output), then server is needed when determining object matching result according to default victory or defeat probability screen Condition, default role's screening conditions and every kind of prediction is selected to combine corresponding first matching probability information and the second matching generally Rate information, filters out P kind prediction combination (P be positive integer) less than or equal to M from the prediction combination of M kind, then from P kind prediction group It is object matching result that a prediction group cooperation is obtained in conjunction.

It should be understood that each player that above-mentioned role's screening conditions can be set as in each troop of prediction combination may make Role guarantees that each player in each troop can select oneself to think according to the selection wish of oneself there is no repetition Role to be used.Correspondingly, the role's prediction result and default role that server is exported according to role's prediction model screen item When part is screened, judge the player for including in each troop of prediction combination in this innings battle the role that may use whether In the presence of repetition, if the hero that equal player may use in each queue does not repeat, it is determined that prediction combination meets the role Screening conditions.

When determining object matching result with specific reference to the first matching probability information and the second matching probability information, server can With elder generation according to the first matching probability information and default victory or defeat probability screening conditions, filter out that meet this pre- from the prediction combination of M kind If the A kind of victory or defeat probability screening conditions predicts combination, further according to the second matching probability information and default role's screening conditions, from A The P kind prediction group of default victory or defeat probability screening conditions and default role's screening conditions is filtered out while met in kind prediction combination It closes；Server can also be filtered out from the prediction combination of M kind first according to the second matching probability information and default role's screening conditions B kind meets the prediction combination of default role's screening conditions, screens item further according to the first matching probability information and default victory or defeat probability Part, filters out from the prediction combination of B kind while the P kind for meeting default victory or defeat probability screening conditions and default role's screening conditions is pre- Survey combination；Server can also consider the first matching probability information and default victory or defeat probability screening conditions and second simultaneously With probabilistic information and default role's screening conditions, disposably from M kind prediction combination in filter out and meanwhile meet preset victory or defeat it is general The prediction combination of the P kind of rate screening conditions and default role's screening conditions.Any restriction is not done to above-mentioned screening sequence herein.

It should be noted that server during executing above-mentioned screening operation, can also further consider second With role's probability of occurrence in probabilistic information, screened according to role's probability of occurrence.Specifically, server may determine that every kind The specific affiliated character types of the player institute role to be used of each troop, then will belong in same troop in prediction combination The corresponding role's probability of occurrence of the role of same character types is added, and is obtained the corresponding probability of occurrence of the character types, is judged angle Whether the corresponding probability of occurrence of color type is more than predetermined probabilities threshold value, if being more than, illustrates that the player in the troop selects this kind The probability of the role of character types is higher, and there is the role for excessively belonging to same character types in the same troop, the team 5 triumph probability may be lower, therefore, can correspondingly screen out this prediction combination.

If should be understood that server according to default victory or defeat probability screening conditions, default role's screening conditions and every kind of prediction Corresponding first matching probability information and the second matching probability information are combined, is filtered out from the prediction combination of M kind a variety of pre- Combination (i.e. above-mentioned P is the integer greater than 1) is surveyed, then server also needs to continue to select a prediction from the prediction combination of this P kind Combination is as object matching as a result, specifically, server can be predicted to select a prediction group cooperation at random in combination from this P kind For object matching as a result, server further can also combine corresponding first matching probability information according to the prediction of this P kind Or the second matching probability information, choose that each battle troop strength is closer or same troop in each player's selection role it is complete The complete higher prediction combination of different probability, as object matching result.

In above-mentioned object matching method provided by the embodiments of the present application, server is based on victory or defeat prediction model to be matched right It as the strength of (such as game player) models, and predicts each prediction and combines corresponding battle victory or defeat probability, be based on each pre- Survey combination it is corresponding battle victory or defeat determine the probability go out object matching as a result, i.e. determination game fighting point side as a result, it is possible to It is effectively prevented from point side for player's capabilities gap great disparity occur, optimization player fights the object matching result of grouping.

It should be understood that whether above-mentioned matching prediction model can accurately determine out the object matching knot for guaranteeing battle balance Fruit, dependent on the model performance of the matching prediction model, and the quality of the model performance of the matching prediction model depends on this Training process with prediction model.According to above-described embodiment it is found that can only include victory or defeat prediction model in matching prediction model, It also may include that victory or defeat prediction model and role's prediction model correspondingly actually may be used to the training process of matching prediction model To be the training process to victory or defeat prediction model, it is also possible to victory or defeat prediction model and training to role's prediction model Journey.

It is first introduced below when matching prediction model only includes victory or defeat prediction model, to the training side of the matching prediction model Method.Referring to Fig. 5, Fig. 5 is a kind of flow diagram of model training method provided by the embodiments of the present application, for ease of description, Following embodiments are described with executing subjects such as servers, it should be appreciated that the model training method can also be applied to other tools The equipment of standby model training function.As shown in figure 5, the model training method the following steps are included:

Step 501: determining the first training sample set, each sample which concentrates includes at least two teams The corresponding at least one attribute information of each object and battle result in 5.

When matching prediction model only includes victory or defeat prediction model, the process of the training matching prediction model is that training should The process of victory or defeat prediction model.It when server training victory or defeat prediction model, needs first to obtain the first training sample set, first instruction Practice and generally include great amount of samples in sample set, it includes participating in certain battle in each sample that corresponding one, each sample, which is fought, Each object (object is player) corresponding at least one attribute information and the battle result of this battle.

When servicing implement body acquisition sample, game fighting information, the game pair can be acquired from game application database War information includes each respective object identity of object and pair of this battle at least two troops for participate in a battle War result；Then, the object identity of server according to each object, acquisition and the corresponding belligerent data of each object, should Supplemental characteristic specifically includes the corresponding belligerent play of each object and battle result；In turn, server can be according to each The respective belligerent data of object determine the corresponding attribute information of each object, such as whole competitive state information and competing in the recent period Skill status information, as the corresponding at least one attribute information of each object；Finally, at least two teams of a battle will be participated in The corresponding at least one attribute information of each object and this fight result as a sample data in 5, so acquire Multiple samples generate the first training sample set.

Specifically, server can be believed the game fighting of any one game of random acquisition from game application database It ceases, it is corresponding right to generally include to participate in each object at least two troops of this game fighting in game fighting information The battle result fought as mark and this.Then, server can be according to the corresponding object identity of each object, from being used for The corresponding belligerent data of each object are obtained in the database of storage object related data, belligerent data herein include pair As all battle plays participated in since establishing the game account and each field fight corresponding battle result.In turn, with Attribute information for whole competitive information and recent competitive information including being illustrated；Server is according to acquired each object Corresponding supplemental characteristic calculates the corresponding whole competitive information of each object and recent competitive information；It is specific to calculate When whole competitive information, server can according to all belligerent plays of the object and each field battle it is corresponding battle as a result, It calculates the whole of the object and fights victory or defeat rate competitive information as a whole；When specifically calculating recent competitive information, server can be with Fought in corresponding battle result from all belligerent plays and each field, in selection nearly one month or it is one week nearly in the object ginseng Corresponding battle is fought as a result, tying in turn according to selected belligerent play and these corresponding battles of battle with these in battlefield time Fruit calculates the recent battle victory or defeat rate of the object as recent competitive information, it should be appreciated that can also choose it according to actual needs The period of his length is as the recent competitive information corresponding period.Finally, will this fight in each object it is respectively right The whole competitive information answered, recent competitive information respectively as the corresponding attribute information of object, and by these attribute informations and The battle result of this battle obtains other sample numbers of the first training sample concentration as sample data in the method According to.

Optionally, the corresponding at least one attribute information of above-mentioned object may include whole competitive information, the object of object Recent competitive information, the ELO ranking system point of object, at least one of the game section of object, the personal information of object； Personal information herein includes the natural quality information of object and/or the game asset information of object.In a upper embodiment Introduction was carried out to the concrete meaning of these attribute informations, referring particularly to the associated description of a upper embodiment, details are not described herein.

It should be understood that server get participate in game fighting the respective object identity of each object after, can basis The respective object identity of each object obtains the ELO ranking of above-mentioned object in the database for storage object related data Any one or more information such as system point, the personal information of the game section of object, object.

It should be understood that in practical applications, server can also obtain sample data by other means, and acquired in utilization Sample data generate the first training sample set, herein not to server obtain sample data specific implementation do any limit It is fixed.

Step 502: each sample training neural network being concentrated according to first training sample, victory or defeat prediction is obtained with training Model, the victory or defeat prediction model are defeated with the corresponding at least one attribute information of each object at least two troops Enter, to fight victory or defeat probability as output.

When server training victory or defeat prediction model, need to construct neural network model in advance as the victory or defeat prediction being trained to The structure of model, the neural network model should be identical as the investment structure of victory or defeat prediction model of practical application, is specifically as follows The structure of victory or defeat prediction model shown in Fig. 3, including deep-neural-network DNN hidden layer, naturally it is also possible to be other structures Neural network model.

It should be understood that the function phase that the function to be realized of the neural network model can be realized with victory or defeat prediction model Together, i.e., the neural network model can be according to the corresponding at least one category of each object at least two troops of input Property information, correspondingly export this and fight corresponding victory or defeat probability.

After neural network model is completed in building, the sample pair of the first training sample concentration obtained in step 501 is utilized The model parameter of the neural network is trained, after the neural network model wait be trained meets the first training termination condition, root According to the model structure and model parameter of the neural network model of the training termination condition of satisfaction first, the victory of building investment practical application Negative prediction model.

When training neural network model, server can be concentrated from the first training sample and be obtained for training the neural network The sample data of model, the corresponding at least one attribute information of each object that a battle is participated in sample data is defeated Enter the neural network model, which passes through to the corresponding at least one of each object for participating in this battle Attribute information is analyzed and processed, and exports the prediction victory or defeat probability of this battle.Server will be true right in the sample data War result is correspondingly converted to true victory or defeat probability, according to the prediction victory or defeat probability of neural network model output and the true victory or defeat Error between probability constructs loss function, in turn, is carried out according to the loss function to the model parameter in neural network model Adjustment, so that the optimization to the neural network model is realized, when neural network model meets the first training termination condition According to the model parameter and model structure of Current Situation of Neural Network model, the victory or defeat prediction model that can put into practical application is generated.

When specifically judging whether above-mentioned neural network model meets the first training termination condition, test sample pair can use First model is verified, which is that the sample concentrated using the first training sample carries out first to neural network model The model that wheel training optimization obtains；Specifically, server respectively corresponds to each object for participating in certain battle in test sample At least one attribute information input first model, it is corresponding to each object of input at least using first model A kind of attribute information is correspondingly handled, and prediction victory or defeat probability is obtained；In turn, according to the true battle result in test sample True victory or defeat probability is generated, is predicted according to the prediction victory or defeat probability calculation victory or defeat that the true victory or defeat probability and the first model export quasi- True rate, when victory or defeat predictablity rate is greater than preset threshold, i.e., it is believed that the model performance of first model can preferably expire Sufficient demand then can generate victory or defeat prediction model according to the model parameter and model structure of first model.

It should be noted that above-mentioned preset threshold can be set according to the actual situation, herein not to the preset threshold It is specifically limited.

Moreover, it is judged that when whether neural network model meets the first training termination condition, it can also be according to through more wheel training Obtained multiple models, it is determined whether continue to be trained model, to obtain the optimal victory or defeat prediction model of model performance.Tool Body, it can use test sample and the multiple neural network models got through more trainings in rotation verified respectively, judge through each Gap between the victory or defeat predictablity rate for the model that training in rotation is got is smaller, then it is assumed that the performance of neural network model has not had There is room for promotion, the highest neural network model of victory or defeat predictablity rate can be chosen, according to the model of the neural network model Parameter and model structure determine victory or defeat prediction model；If the victory or defeat prediction of the neural network model got through each training in rotation is accurate There is biggish gap, then it is assumed that the performance of the neural network model can continue there are also the space promoted to the nerve between rate Network model is trained, until obtaining the most stable and optimal victory or defeat prediction model of model performance.

Step 503: matching prediction model is determined according to the victory or defeat prediction model.

Training obtain meet first training termination condition victory or defeat prediction model after, directly using the victory or defeat prediction model as Prediction model is matched, is respectively corresponded to using the matching prediction model according to each object to be matched in the various predictions of input combination At least one attribute information, determine that corresponding victory or defeat probability is combined in various predictions.

Using above-mentioned matching prediction model training method, the sample concentrated using the first training sample is to the mind constructed in advance It is trained through network model, the sample which concentrates respectively is corresponded to including each object at least two troops At least one attribute information and battle result；Victory is finally generated according to the neural network model of the training termination condition of satisfaction first Negative prediction model, and using the victory or defeat prediction model as matching prediction model.In the training process, based at least two troops The corresponding at least one attribute information of each object is trained neural network model, can guarantee the neural network mould Type more accurately learns the feature for being able to reflect object strength out according to the attribute information of input, in turn, based on this, more quasi- Really predict that these objects participate in the victory or defeat probability of same field battle.

Next it introduces when matching prediction model includes victory or defeat prediction model and role's prediction model, which is predicted The training method of model.It should be noted that the training method of the victory or defeat prediction model in the matching prediction model and above-mentioned Fig. 5 Shown in training method it is identical, therefore, will mainly introduce the training method of role's prediction model in the following embodiments.

Referring to Fig. 6, Fig. 6 is the flow diagram of another model training method provided by the embodiments of the present application, in order to just In description, following embodiments are described with executing subjects such as servers, it should be appreciated that the model training method can also be applied to Other have the equipment of model training function.As shown in fig. 6, method includes the following steps:

Step 601: determining the second training sample set, each sample which concentrates includes that object participates in T+ Role's sequence used in 1 innings of battle (T is the positive integer more than or equal to 1).

When server training role's prediction model, need first to obtain the second training sample set, which concentrates Great amount of samples is generally included, each sample corresponds to an object, includes that object T+1 innings of battles of participation are used in each sample Role's sequence.

When servicing implement body acquisition sample, the object identity of some object can be obtained at random, and then according to the object Object identity, from obtained in the database for storage object related data the object continuous T+1 innings battle used in Role, according to this T+1 innings battle used in role creation role's sequence.So obtain multiple objects participate in T+1 innings it is right Role's sequence used in fighting, and then generate the second training sample set.

It should be understood that above-mentioned T be it is any be more than or equal to 1 positive integer, i.e., server need to obtain object more innings battle in Used role's sequence, is usually set as 10 or 20 for T, certainly, in practical applications, can also be according to actual needs by T It is set as other numerical value, any restriction is not done to the specific value of T herein.

It should be understood that in practical applications, server can also obtain sample by other means, and utilize acquired sample The second training sample set of this generation, the specific implementation for not obtaining sample to server herein do any restriction.

Step 602: each sample training neural network being concentrated according to second training sample, role's prediction is obtained with training Model, which fights used role sequence column at first T innings with object as input, and participates in T+1 with object Office fights used role and role's probability of occurrence is output.

When server training role's prediction model, need to construct neural network model in advance as the role's prediction being trained to The structure of model, the neural network model should be identical as the investment structure of role's prediction model of practical application, is specifically as follows The structure of role's prediction model shown in Fig. 4, it is hidden including deep-neural-network DNN hidden layer and shot and long term memory network LSTM Layer, naturally it is also possible to be the neural network model of other structures.

It should be understood that the function phase that the function to be realized of the neural network model can be realized with role's prediction model Together, i.e. the role's sequence that can be used according to the object of input at first T innings of the neural network model, correspondingly exports the object and exists The T+1 innings of roles used and role's probability of occurrence.

After neural network model is completed in building, the sample pair of the second training sample concentration obtained in step 601 is utilized The model parameter of the neural network is trained, after the neural network model wait be trained meets the second training termination condition, root According to the model structure and model parameter of the neural network model of the training termination condition of satisfaction second, the angle of building investment practical application Color prediction model.

When training neural network model, server can be concentrated from the second training sample and be obtained for training the neural network The sample data of model, by the role's sequence inputting neural network model used in first T innings battle of object in sample data, The neural network model exports object at T+1 innings by being analyzed and processed to role's sequence used in first T innings battle Predict that the role used and role predict probability of occurrence in battle.Server fights object in the sample data at T+1 innings In the role that really uses correspondingly be converted to the true probability of occurrence of role, predicted according to the role that neural network model exports Error constructs loss function between existing probability and the true probability of occurrence of role, in turn, according to the loss function to neural network mould Model parameter in type is adjusted, so that the optimization to the neural network model is realized, when neural network model meets second When training termination condition, reality can be put into according to the model parameter and model structure of Current Situation of Neural Network model, generation Role's prediction model of application.

When specifically judging whether above-mentioned neural network model meets the second training termination condition, test sample pair can use First model is verified, which is that the sample concentrated using the second training sample carries out first to neural network model The model that wheel training optimization obtains；Specifically, server by object in test sample first T innings battle used in role's sequence First model is inputted, is carried out using object first T innings battle used in role sequence of first model to input corresponding Ground processing obtains fighting the role for predicting to use in T+1 innings of battles and role predicts probability of occurrence；In turn, according to test The true probability of occurrence of role creation role that object really uses in T+1 innings in sample really occurs general according to the role There is probability calculation role's predictablity rate in role's prediction of rate and the output of the first model, presets when role's predictablity rate is greater than It, i.e., then can be according to first model it is believed that the model performance of first model preferably can satisfy demand when threshold value Model parameter and model structure generate role's prediction model.

Moreover, it is judged that when whether neural network model meets the second training termination condition, it can also be according to through more wheel training Obtained multiple models, it is determined whether continue to be trained model, to obtain the optimal role's prediction model of model performance.Tool Body, it can use test sample and the multiple neural network models got through more trainings in rotation verified respectively, judge through each Gap between the role's predictablity rate for the model that training in rotation is got is smaller, then it is assumed that the performance of neural network model has not had There is room for promotion, the highest neural network model of role's predictablity rate can be chosen, according to the model of the neural network model Parameter and model structure determine role's prediction model；If role's prediction of the neural network model got through each training in rotation is accurate There is biggish gap, then it is assumed that the performance of the neural network model can continue there are also the space promoted to the nerve between rate Network model is trained, until obtaining the most stable and optimal role's prediction model of model performance.

Step 603: matching prediction model, the matching prediction model are determined according to victory or defeat prediction model and role's prediction model Including victory or defeat prediction model and role's prediction model.

Training obtains after meeting role's prediction model of the second training termination condition, it is terminated item with the training of satisfaction first The victory or defeat prediction model of part combines composition matching prediction model, using the matching prediction model according to the various predictions of input In combination in the corresponding at least one attribute information of each object and prediction combination each object fought at first T innings in make Role's sequence determines that various predictions combine corresponding victory or defeat probability and predict that each object is at T+1 innings in combination In role's prediction result.

It should be understood that the victory or defeat prediction model and role's prediction model in the matching prediction model can be independent two minds Through network, its corresponding function is respectively realized；Certainly, the victory or defeat prediction model and role's prediction model in the matching prediction model It may be the same neural network, i.e., the neural network can be achieved at the same time victory or defeat prediction model and role's prediction model institute energy The function of enough realizing.

Using above-mentioned matching prediction model training method, the sample concentrated using the second training sample is to the mind constructed in advance It is trained through network model, the sample which concentrates includes object role sequence used in T+1 innings of battles Column；Role's prediction model is generated according to the neural network model of the training termination condition of satisfaction second, and mould is predicted according to the role Type and victory or defeat prediction model generate matching prediction model.The matching prediction model can predict that various prediction combinations are corresponding simultaneously Each object role used in this battle in victory or defeat probability and prediction combination, consequently facilitating server is according to predicting Victory or defeat probability and each object role used in this battle, determine final object matching as a result, guaranteeing the target Worthy opponent with result Zhong Ge troop, and guarantee that each player for participating in battle can select role according to itself wish.

The application is implemented below with reference to Fig. 7 for a further understanding of object matching method provided by the embodiments of the present application Example provides object matching method and carries out whole introduction.

Referring to Fig. 7, Fig. 7 is the overall work configuration diagram of object matching method provided by the embodiments of the present application.Such as Fig. 7 Shown, this method includes off-line training step and application on site stage.

In off-line training step, each player that server obtains the same field battle of participation from player's offline database is each Self-corresponding at least one attribute information, and the battle of this battle is obtained as a result, each player is corresponding at least A kind of sample data that attribute information is concentrated with battle result as the first training sample, the sample for utilizing the first sample to concentrate Data are trained neural network model, obtain victory or defeat prediction model；In addition, server can also be from player's offline database Middle some player of acquisition role's sequence used in T+1 innings of battles is utilized as the sample data that the second training sample is concentrated The sample data that second training sample is concentrated is trained neural network model, obtains role's prediction model.

It should be noted that above-mentioned at least one attribute information can specifically include player whole competitive state information, The recent competitive state information of player, the ELO ranking system point of player, the game section of player, player personal information in extremely Few one kind；Wherein, the personal information of player includes the natural quality information of player and/or the game asset information of player.

In the application on site stage, server is after receiving the battle request of client initiation, according to the ELO of player points And/or the game section of player, it is chosen from all players for initiating battle request in the same period and participates in what same field was fought Player will participate in all players of this battle as object set to be matched, and according to the corresponding player identification of each player, The corresponding at least one attribute information of each player is obtained from player's offline database and each player is right at first T innings Role's sequence used in war is added in the object set to be matched；In turn, according to the object set to be matched, group is utilized It closes number calculation formula and generates the prediction combination of M kind.

By the corresponding at least one attribute information of each player and each player in every kind of prediction combination T innings first Role's sequence used in battle is input to matching prediction model, and the victory or defeat prediction model matched in prediction model is correspondingly predicted Corresponding battle victory or defeat probability is combined in every kind of prediction, and role's prediction model in matching probability model correspondingly predicts each player The role that may be used in this battle, i.e. role's prediction result.

In turn, server combines corresponding battle victory or defeat probability and player according to every kind of prediction that matching probability model exports The role that may use in this battle determines object matching as a result, determining that game is distinguished as a result, and by the object matching As a result it is correspondingly back to client, so that player knows game respectively as a result, and participating in game fighting.

For above-described object matching method, present invention also provides corresponding object matching devices, so that above-mentioned The application and realization of object matching method in practice.

It is a kind of knot of object matching device 800 corresponding with object matching method shown in figure 2 above referring to Fig. 8, Fig. 8 Structure schematic diagram, the object ticket device 800 include:

Obtain module 801, include for obtaining object set to be matched, in the object set to be matched at least two to Object is matched, the object to be matched includes at least one attribute information；

Prediction combination determining module 802, for generating the prediction combination of M kind according to the object set to be matched, M is positive whole Number, wherein include at least two queues in every kind of prediction combination, include at least an object to be matched in any queue；

Prediction module 803, for obtaining the corresponding matching probability letter of every kind of prediction combination by matching prediction model Breath；Wherein, the matching prediction model includes victory or defeat prediction model, and the victory or defeat prediction model is with each at least two queues The corresponding at least one attribute information of object is input, to fight victory or defeat probability as output；

Determining module 804 determines object matching knot for combining corresponding matching probability information according to every kind of prediction Fruit.

Optionally, on the basis of object matching device shown in Fig. 8, the matching prediction model further includes role's prediction Model, role's prediction model fights used role sequence column at first T innings with object as input, and participates in T+ with object Role used in 1 innings of battle and role's probability of occurrence are output；

Then the prediction module 803 is specifically used for:

By the victory or defeat prediction model in matching prediction model, obtains every kind of prediction and combines corresponding victory or defeat probability, As the first matching probability information；

By role's prediction model in matching prediction model, obtain every kind of prediction combine in the corresponding role of each object Prediction result, as the second matching probability information；Wherein, role's prediction result makes in T+1 innings for characterizing object Role and role's probability of occurrence；

Then the determining module 804 is specifically used for:

It is respectively corresponding according to default victory or defeat probability screening conditions, default role's screening conditions and every kind of prediction combination The first matching probability information and the second matching probability information, from the M kind prediction combination in select P kind prediction combination, wherein institute Stating P is the positive integer less than or equal to the M；

A prediction combination is obtained from P kind prediction combination, as object matching result.

Optionally, on the basis of object matching device shown in Fig. 8, the acquisition module 801 is specifically used for:

Receive the matching request that game application server is sent, include in the matching request 2N objects to be matched respectively Corresponding at least one attribute information, the 2N objects to be matched are game section and/or angstrom Lip river ranking system according to object System point determination, wherein N is positive integer；

Obtain object set to be matched from the matching request, include in the object set to be matched the 2N to Match object.

Optionally, on the basis of object matching device shown in Fig. 8, the prediction combination determining module 802 is specifically used In:

Determine the object total number of object to be matched in the object set to be matched；

According to membership in the object total number and default team, determine that M kind is predicted by number of combinations calculation formula Combination, the M are number of combinations.

Optionally, on the basis of object matching device shown in Fig. 8, at least one attribute information includes: object Whole competitive state information, the recent competitive state information of object, the game section of angstrom Lip river ranking system point of object, object At least one of position, the personal information of object；Wherein, the personal information includes the natural quality information of object and/or right The game asset information of elephant.

In above-mentioned object matching device provided by the embodiments of the present application, server is based on victory or defeat prediction model to be matched right As the strength of (such as game player) is modeled to predict prediction and combine corresponding battle victory or defeat probability, and based on each prediction It combines corresponding victory or defeat determine the probability and goes out object matching as a result, determining point side of game fighting as a result, it is possible to be effectively prevented from There is the side of dividing of player's capabilities gap great disparity, optimization player fights the matching result of grouping.

For above-described model training method, the embodiment of the present application also provides corresponding model training apparatus, with Convenient for the application and realization of above-mentioned model training method in practice.

It is a kind of knot of model training apparatus 900 corresponding with model training method shown in figure 5 above referring to Fig. 9, Fig. 9 Structure schematic diagram, the model training apparatus 900 include:

Training sample determining module 901, for determining each of the first training sample set, the first training sample concentration Sample includes the corresponding at least one attribute information of each object at least two troops and battle result；

Training module 902, for concentrating each sample training neural network according to first training sample, with trained To victory or defeat prediction model, the victory or defeat prediction model is with the corresponding at least one category of each object at least two troops Property information be input, with fight victory or defeat probability for output；

Model determining module 903, for determining matching prediction model according to the victory or defeat prediction model.

Optionally, on the basis of above-mentioned model training apparatus shown in Fig. 9, the training sample determining module 901, also For determining that the second training sample set, each sample that second training sample is concentrated include that object participates in T+1 innings of battle institutes The role's sequence used, T are the positive integer more than or equal to 1；

The training module 902 is also used to concentrate each sample training neural network according to second training sample, with Training obtains role's prediction model, and it is input that role's prediction model, which fights used role sequence column at first T innings with object, And used role and role's probability of occurrence are fought as output with T+1 innings of object participation；

Then the model determining module 903 is specifically used for:

Matching prediction model is determined according to the victory or defeat prediction model and role's prediction model, and mould is predicted in the matching Type includes the victory or defeat prediction model and role's prediction model.

Optionally, on the basis of above-mentioned model training apparatus shown in Fig. 9, the training sample determining module 901 has Body is used for:

Game fighting information is acquired from game application database, the game fighting information includes participating in a battle Each respective object identity of object and the battle result of this battle at least two troops；

Object identity according to each object obtains and the corresponding belligerent data of each object, the belligerent data packet Belligerent play and battle are included as a result, respective belligerent data determine the corresponding at least one of each object according to each object Attribute information；

To participate in the corresponding at least one attribute information of each object at least two troops that one is fought and be somebody's turn to do The battle result of field battle is a sample data, acquires at least one sample data and generates the first training sample set.

Optionally, on the basis of above-mentioned model training apparatus shown in Fig. 9, at least one attribute information includes: Whole competitive state information, recent competitive state information, angstrom Lip river ranking system point of object, the game section of object, object At least one of personal information；The personal information includes the natural quality information of object and/or the game asset letter of object Breath.

Above-mentioned model training apparatus using the first training sample concentrate sample to the neural network model constructed in advance into Row training, the sample which concentrates include the corresponding at least one attribute letter of each object for participating in battle Breath and battle result；Victory or defeat prediction model is finally generated according to the neural network model of the training termination condition of satisfaction first, and will The victory or defeat prediction model is as matching prediction model.In the training process, corresponding based on each object for participating in battle At least one attribute information is trained neural network model, and the neural network model can be guaranteed according to these attribute informations More accurately learn the feature for being able to reflect object strength out, in turn, based on this, more accurately predicts that these objects participate in The victory or defeat probability of same field battle.

Present invention also provides a kind of equipment for matching object, which is specifically as follows server, referring to Figure 10, Figure 10 be it is provided by the embodiments of the present application a kind of for determining the server architecture schematic diagram of matching result, which can Bigger difference is generated because configuration or performance are different, may include one or more central processing units (central Processing units, CPU) 1022 (for example, one or more processors) and memory 1032, one or one with The storage medium 1030 (such as one or more mass memory units) of upper storage application program 1042 or data 1044.Its In, memory 1032 and storage medium 1030 can be of short duration storage or persistent storage.It is stored in the program of storage medium 1030 It may include one or more modules (diagram does not mark), each module may include to the series of instructions in server Operation.Further, central processing unit 1022 can be set to communicate with storage medium 1030, execute on server 1000 Series of instructions operation in storage medium 1030.

Server 1000 can also include one or more power supplys 1026, one or more wired or wireless nets Network interface 1050, one or more input/output interfaces 1058, and/or, one or more operating systems 1041, example Such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM etc..

The step as performed by server can be based on the server architecture shown in Fig. 10 in above-described embodiment.

In the server for determining matching result, CPU 1022 is for executing following steps:

Optionally, any specific implementation of object matching method in the embodiment of the present application can also be performed in CPU1022 Method and step.

In addition, the equipment is specifically as follows server, the clothes present invention also provides a kind of equipment for training pattern Be engaged in device structure with shown in Fig. 10 for determining that the structure of server of matching result is similar, CPU therein for execution with Lower step:

Optionally, any specific implementation of model training method provided by the embodiments of the present application can also be performed in CPU Method and step.

The embodiment of the present application also provides a kind of computer readable storage medium, for storing program code, the program code For executing any one embodiment or a kind of model in a kind of object matching method described in foregoing individual embodiments Any one embodiment in training method.

The embodiment of the present application also provides a kind of computer program product including instruction, when run on a computer, So that computer executes any one embodiment in a kind of object matching method described in foregoing individual embodiments, Huo Zheyi Any one embodiment in kind model training method.

It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description, The specific work process of device and unit, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.

In several embodiments provided herein, it should be understood that disclosed system, device and method can be with It realizes by another way.For example, the apparatus embodiments described above are merely exemplary, for example, the unit It divides, only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components It can be combined or can be integrated into another system, or some features can be ignored or not executed.Another point, it is shown or The mutual coupling, direct-coupling or communication connection discussed can be through some interfaces, the indirect coupling of device or unit It closes or communicates to connect, can be electrical property, mechanical or other forms.

The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme 's.

It, can also be in addition, each functional unit in each embodiment of the application can integrate in one processing unit It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list Member both can take the form of hardware realization, can also realize in the form of software functional units.

If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product When, it can store in a computer readable storage medium.Based on this understanding, the technical solution of the application is substantially The all or part of the part that contributes to existing technology or the technical solution can be in the form of software products in other words It embodies, which is stored in a storage medium, including some instructions are used so that a computer Equipment (can be personal computer, server or the network equipment etc.) executes the complete of each embodiment the method for the application Portion or part steps.And storage medium above-mentioned includes: USB flash disk, mobile hard disk, read-only memory (full name in English: Read-Only Memory, english abbreviation: ROM), random access memory (full name in English: Random Access Memory, english abbreviation: RAM), the various media that can store program code such as magnetic or disk.

The above, above embodiments are only to illustrate the technical solution of the application, rather than its limitations；Although referring to before Embodiment is stated the application is described in detail, those skilled in the art should understand that: it still can be to preceding Technical solution documented by each embodiment is stated to modify or equivalent replacement of some of the technical features；And these It modifies or replaces, the spirit and scope of each embodiment technical solution of the application that it does not separate the essence of the corresponding technical solution.

Claims

1. a kind of object matching method characterized by comprising

Object set to be matched is obtained, includes at least two objects to be matched in the object set to be matched, it is described to be matched Object includes at least one attribute information；

The prediction combination of M kind is generated according to the object set to be matched, M is positive integer, wherein include extremely in every kind of prediction combination Lack two queues, includes at least an object to be matched in any queue；

Every kind of corresponding matching probability information of prediction combination is obtained by matching prediction model；Wherein, the matching prediction Model includes victory or defeat prediction model, and the victory or defeat prediction model is with each object at least two queues corresponding at least one Attribute information is input, to fight victory or defeat probability as output；

2. method according to claim 1, which is characterized in that the matching prediction model further includes role's prediction model, institute Role's prediction model is stated with object role sequence column used in first T innings battle as input, and T+1 innings of battles are participated in object Used role and role's probability of occurrence are output；

It is then described to obtain every kind of corresponding matching probability information of prediction combination by matching prediction model, comprising:

By the victory or defeat prediction model in matching prediction model, obtains every kind of prediction and combine corresponding victory or defeat probability, as First matching probability information；

By role's prediction model in matching prediction model, obtains the corresponding role of each object in every kind of prediction combination and predict As a result, as the second matching probability information；Wherein, role's prediction result is for characterizing object used in T+1 innings Role and role's probability of occurrence；

It is then described that corresponding matching probability information is combined according to every kind of prediction, determine object matching result, comprising:

Corresponding the is combined according to default victory or defeat probability screening conditions, default role's screening conditions and every kind of prediction One matching probability information and the second matching probability information select the prediction combination of P kind, wherein the P from M kind prediction combination For the positive integer less than or equal to the M；

3. method according to claim 1, which is characterized in that described to obtain object set to be matched, comprising:

The matching request that game application server is sent is received, includes that 2N objects to be matched respectively correspond in the matching request At least one attribute information, the 2N objects to be matched are according to the game section of object and/or angstrom Lip river ranking system point Determining, wherein N is positive integer；

Object set to be matched is obtained from the matching request, includes the 2N to be matched in the object set to be matched Object.

4. method according to claim 1, which is characterized in that described to generate the prediction of M kind according to the object set to be matched Combination, comprising:

According to membership in the object total number and default team, the prediction combination of M kind is determined by number of combinations calculation formula, The M is number of combinations.

5. method according to claim 1, which is characterized in that at least one attribute information includes: that the entirety of object is competing Skill status information, the recent competitive state information of object, angstrom Lip river ranking system point of object, the game section of object, object At least one of personal information；Wherein, the personal information includes the natural quality information of object and/or the game money of object Produce information.

6. a kind of model training method characterized by comprising

Determine that the first training sample set, each sample that first training sample is concentrated include each right at least two troops As corresponding at least one attribute information and battle result；

Each sample training neural network is concentrated according to first training sample, victory or defeat prediction model is obtained with training, it is described Victory or defeat prediction model is input with the corresponding at least one attribute information of each object at least two troops, with battle Victory or defeat probability is output；

7. method according to claim 6, which is characterized in that the method also includes:

Determine that the second training sample set, each sample that second training sample is concentrated include that object participates in T+1 innings of battle institutes The role's sequence used, T are the positive integer more than or equal to 1；

Each sample training neural network is concentrated according to second training sample, role's prediction model is obtained with training, it is described Role's prediction model fights used role sequence column at first T innings with object as input, and participates in T+1 innings of battle institutes with object The role used and role's probability of occurrence are output；

It is then described that matching prediction model is determined according to the victory or defeat prediction model, comprising:

Matching prediction model, the matching prediction model packet are determined according to the victory or defeat prediction model and role's prediction model Include the victory or defeat prediction model and role's prediction model.

8. according to claim 6 or 7 the method, which is characterized in that first training sample set of determination, comprising:

Game fighting information is acquired from game application database, the game fighting information includes participating in a battle at least Each respective object identity of object and the battle result of this battle in Liang Ge troop；

Object identity according to each object, which is obtained, includes ginseng with the corresponding belligerent data of each object, the belligerent data Battlefield time and battle are as a result, respective belligerent data determine the corresponding at least one attribute of each object according to each object Information；

The corresponding at least one attribute information of each object and this are right at least two troops to participate in a battle The battle result of war is a sample data, acquires at least one sample data and generates the first training sample set.

9. according to claim 6 or 7 the method, which is characterized in that at least one attribute information includes: the entirety of object Competitive state information, the recent competitive state information of object, angstrom Lip river ranking system point of object, object game section, object At least one of personal information；The personal information includes the natural quality information of object and/or the game asset of object Information.

10. a kind of object matching device, which is characterized in that described device includes:

Module is obtained, includes at least two to be matched right in the object set to be matched for obtaining object set to be matched As the object to be matched includes at least one attribute information；Prediction combination determining module, for according to the object to be matched Set generates the prediction combination of M kind, and M is positive integer, wherein including at least two queues in every kind of prediction combination, in any queue Including at least an object to be matched；

Prediction module, for obtaining every kind of corresponding matching probability information of prediction combination by matching prediction model；Wherein, The matching prediction model includes victory or defeat prediction model, the victory or defeat prediction model with each object at least two queues respectively Corresponding at least one attribute information is input, to fight victory or defeat probability as output；

11. a kind of model training apparatus, which is characterized in that described device includes:

Training sample determining module, for determining the first training sample set, each sample packet that first training sample is concentrated Include the corresponding at least one attribute information of each object and the battle result at least two troops；

Training module obtains victory or defeat for concentrating each sample training neural network according to first training sample with training Prediction model, the victory or defeat prediction model is with the corresponding at least one attribute information of each object at least two troops For input, exported with fighting victory or defeat probability；

12. a kind of server, which is characterized in that the server includes processor and memory:

The processor is used for according to the instruction execution object described in any one of claim 1 to 5 in said program code Method of completing the square.

13. a kind of server, which is characterized in that the server includes including processor and memory:

The processor is used to be instructed according to the described in any item models of instruction execution claim 6 to 9 in said program code Practice method.

14. a kind of computer readable storage medium, including instruction, when run on a computer, so that computer executes such as Object matching method described in any one of claims 1 to 5, or execute the mould as described in any one of claim 6 to 9 Type training method.