KR20050096791A - Gamer's game style transplanting system and its processing method by artificial intelligence learning - Google Patents

Abstract

The present invention is a game style transplantation system and a method of transplanting the same, more specifically, a game style such as a gamer's way of playing a game or a habit, and a variety of game characters by applying the learned game style to the NPC of the game. The present invention relates to a game style porting system and a method of porting the same that enhance the fun of the game.

The game style transplantation system according to the present invention defines the progress of the game as the relationship between the style vector, the weight vector, and the output vector, and learns the weight vector to produce similar game output according to the game style through the game log of test gamers. Learn the style vector from the game output obtained through the game log of the target gamer, infer the output vector through the matrix operation of the learned style vector and the weight vector, and match the inferred output vector. Port the output to NPC.

According to the present invention, an artificial intelligence NPC that learns a game style of a specific gamer through the provision of a game style transplantation system and a transplantation method is active in the game, thereby making it possible to construct a more dynamic game through a more intelligent game character. .

Description

Gamer's game style transplanting system and its processing method by artificial intelligence learning}

The present invention relates to a game style transplantation system and a transplantation method of the gamers through artificial intelligence learning, and more particularly, to learn game styles such as the manner or habit of the gamers playing the game, and to learn the game style of the game. The present invention relates to a game style porting system and a method of porting the same, which provides a variety of game characters and enhances the fun of the game by applying it to a character.

In the early games, the game progress was simple, and when the gamers repeated the game, the game pattern was found to be repeated, the gamers learned the repetitive rules, predicted the progress of the game, and formulated game manipulation in the prediction situation. Played the game.

In recent years, the development of online games, mud games, RPG games, and the like, breaking away from the aforementioned monotonous game progress and providing various game characters and game situations, more dynamic games are being introduced.

More dynamic games require the control of a variety of characters and game situations, which are effectively handled by artificial intelligence processing techniques that perform highly complex decision making and present highly optimized solutions. Is implemented. In addition, beyond the traditional computer-to-gamer and gamer-to-gamer form, AI plays games with players on behalf of the computer, or NPCs (Agent NPCs) playing games with other players on behalf of the owner. The intelligent game of technology is evolving into a futuristic game.

As a prior art, the published patent (JP-A-2003-72640) discloses a game system and a game method in which a cyber clone, which has learned the propensity of ownership, is active. However, the published patent refers to the pure probability of each input in a specific situation in which the cyber clone learns the propensity of ownership, which is not analyzed by the style element of the owner.

This is because the learning is performed statistically through the pure input manipulation of the gamers without the heuristic information such as the game propensity and the reason why such a game manipulation is performed in a specific situation, and thus the validity of the learning result may be lowered and should be dealt with. There is a problem that it is a burden on the game system process because the load of data and processing to do has an exponentially increasing complexity.

The present invention was devised to solve the above problems, and after learning the weight vector representing the correlation between the game style and the game output of the gamers through the game log of the test gamers, and inversely from the game outputs of the transplant target gamers It is an object of the present invention to provide a game style transplantation system and a method of transplanting the game style through the AI learning that the agent NPC supports the game based on the game style learned from the specific gamer by inferring and learning the user's game style.

In addition, by using empirical information such as game style information of gamers, weight information for inducing a specific game output in a specific situation, and game output information determined by a formula between game style and weight information, the game of the actual gamers is used. It is aimed to provide a game style porting system and a method of porting the game, which predicts a game style with high similarity to the operation, and the processing process puts less load on the process.

Game style transplantation system of gamers through artificial intelligence learning according to the present invention in order to achieve the above object, in the game style transplantation system to learn the game style of the target gamers transplanted to the NPC, style vector, game state A vector definition module for storing items of a weight vector and an output vector in a vector definition DB; A game log generation module for generating a game log by informing a process of performing a game by a plurality of test gamers and storing the information in a game log DB; A style vector digitizing module for analyzing and digitizing game styles of individual test gamers stored in the game log DB and storing them in a style vector DB; A weight vector learning module for learning the weight vector for each general game state by storing the weight vector learned from the current weight vector in the weight vector DB based on the game output of the test gamers stored in the game log DB; An individual gamer style vector learning module for learning the inferred style vector by storing the style vector derived based on the game log of the game player to be transplanted into the style vector DB; A game output inference module for predicting and determining an output vector from a style vector learned by the individual gamer style vector learning module and a weight vector of a current game state for a player to be transplanted; And dependent on a specific game operating system, receiving a current game state from the game operating system, and matching the output vector generated by the game output inference module in order to port the game style of the specific gamer to the NPC of the game operating system. It characterized in that it comprises a game adapter module for delivering a game output to the game operating system.

According to a preferred feature of the present invention, the vector definition module defines a style item that defines psychological information, habits, and abilities when the gamer progresses the game as a style vector, and is specified by the game style vector for each specific game state. By defining a weight vector consisting of weights representing the associations resulting from the output vector, and defining game output items as output vectors that gamers can output as game actions through manipulation of input devices in specific game situations, It can be applied to various games independently of its kind.

In addition, the game state is determined by the number of total game states according to a state component defining a state in which a game is played and the number of divisions given per unit of the state component, so that the state component and the number of divisions are large. It is characterized by forming a more detailed game state.

In the present invention, the weight vector learning module includes: a game state definition DB in which a game state identification code matching the game state defined by the division number for each state component is stored; A game state acquisition module for obtaining a game state identification code corresponding to a game state of a test gamer read from the game log DB from the game state definition DB; A game output definition DB that contains a game output identification code that matches a game output by a user's input operation in a specific game state; A game output acquisition module for obtaining a game output identification code corresponding to a game output of a test gamer read from the game log DB from the game output definition DB; A style vector DB for storing a style vector of the individual test gamers analyzed by the style vector digitizing module; A style vector acquisition module for obtaining a style vector from the style vector DB; A weight vector DB for storing a weight vector for each game state for mapping the style vector to an output vector for each specific game state; And obtaining a current weight vector corresponding to a game state obtained by the game state obtaining module from the weight vector DB, and generating a new weight vector through calculation of the current weight vector and the style vector obtained by the style vector obtaining module. And a weight calculation processing module for repeating learning by storing the data in the weight vector DB again, and calculating the weight vector through calculation of the weight calculation module based on game information of test gamers read from the game log DB. It is characterized by learning repeatedly.

In the present invention, the individual gamer style vector learning module, game state definition DB that contains a game state identification code matching the game state defined by the number of divisions by state component; A game state acquisition module for acquiring a game state identification code corresponding to a game state of a game player to be transplanted read from the game log DB from the game state definition DB; A game output definition DB that contains a game output identification code that matches a game output by a user's input operation in a specific game state; A game output acquisition module for obtaining a game output identification code corresponding to a game output of a test gamer read from the game log DB from the game output definition DB; A style vector DB for storing a style vector of the game player to be transplanted; A style vector acquisition module for obtaining a current style vector from the style vector DB; A weight vector DB for storing a weight vector for each game state for mapping the style vector to an output vector for each specific game state; And obtaining a weight vector corresponding to a game state obtained by the game state obtaining module from the weight vector DB, and calculating the weight vector and the current style vector of the transplant target gamer acquired by the style vector obtaining module. And a style operation processing module for generating a vector and storing the same in the style vector DB and repeating the learning. The operation of the style calculation processing module is performed based on game information of a game player to be transplanted read from the game log DB. Iteratively performing a style vector, and by converging the style vector value of the gamer to be transplanted from the patterned weight vector value, characterized in that learning.

In the present invention, in order to achieve the above object, the game style transplantation method of the gamers through the artificial intelligence learning according to the present invention, in the game style transplantation method to learn the game style of the target gamers transplanted to NPC, (1) defining items of a style vector, a weight vector for each game state, and an output vector and storing the items in a vector definition DB; (2) generating a game log by storing information in a game log DB of a plurality of test gamers repeatedly performing a game; (3) analyzing and quantifying the game styles of the individual test gamers stored in the game log DB and storing them in the style vector DB; (4) learning a weight vector causing a specific output vector in a specific game state by style vectors of a plurality of test gamers; (5) learning a style vector of the gamer according to a game output inputted by the gamer from the game log of the game player to be transplanted; And (6) receiving a game state from a game operating system, inferring and determining an output vector based on a learned style vector of a game player to be transplanted, and delivering an output value corresponding to the output vector to a game operating system to determine a specific gamer's ability. And porting the game style to the NPC.

In the present invention, the step (4) comprises the steps of: (4-1) setting an initial value to the learning environment variable; (4-2) setting an arbitrary initial value in the weight vector DB; (4-3) reading game log information of a test gamer from the game log DB; (4-4) obtaining a game state from the game log information; (4-5) reading the current weight vector according to the game state from the weight vector DB; (4-6) reading the style vector of the gamer from the gamer style DB constructed in the step (3); And (4-7) generating a new weight vector through operation from the current weight vector of step (4-5) and the style vector of step (4-6), and storing the new weight vector again in the weight vector DB. And repeatedly learning the weight vector through calculation based on game information of test gamers read from the game log DB.

In the present invention, the step (5), (5-1) setting the initial value to the learning environment variable; (5-2) setting the game style vector of the gamer to be transplanted to the style vector DB as an initial value; (5-3) reading game log information of a test gamer from the game log DB; (5-4) obtaining a game state from the game log information; (5-5) reading the weight vector according to the game state from the weight vector DB; (5-6) reading the current style vector of the gamer from the gamer style DB constructed in the step (5-2); And (5-7) generating a new style vector through operation from the weight vector of step (5-5) and the current style vector of step (5-6), and storing the new style vector again in the style vector DB for learning. It is characterized in that the repetitive learning of the style vector through the operation based on the game information of the target gamers to be read from the game log DB.

In the present invention, the step (6), (6-1) receiving the game state that the NPC in the game operating system; (6-2) reading the style vector of the player to be transplanted learned in the step (5) from the style vector DB; (6-3) reading the weight vector according to the game state of step (6-1) from the weight vector DB learned in step (4); (6-4) generating an output vector by computing the style vector of step 6-2 and the weight vector of step 6-3; And (6-5) delivering the game output corresponding to the vector value corresponding to the highest value among the output vectors to the game operating system, wherein the style vector and the test for learning the style of the game object to be transplanted are tested. It is characterized in that the output is predicted through operation with the weight vector learned from the game log of the gamers and transferred to the game operating system and transplanted into the NPC.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms or words used in the specification and claims should not be construed as having a conventional or dictionary meaning, and the inventors should properly explain the concept of terms in order to best explain their own invention. Based on the principle that can be defined, it should be interpreted as meaning and concept corresponding to the technical idea of the present invention. Therefore, the embodiments described in the specification and the drawings shown in the drawings are only the most preferred embodiment of the present invention and do not represent all of the technical idea of the present invention, various modifications that can be replaced at the time of the present application It should be understood that there may be equivalents and variations.

1 illustrates the concept of a learning environment that is modeled after the game operation of a gamer in the technical configuration according to the present invention. In the present invention is to implement the NPC transplanted style according to the game operation style of the game target player. To do this, the game operation method of the gamers in the human world is used as a model for learning and prediction, and the operation of the game is analyzed by converting it into a computer environment model using artificial intelligence technology.

In the configuration and the embodiment of the present invention, for convenience of explanation, an action-related game (eg, a street fighter) between a gamer and a computer-operated non player character (NPC) is described. NPCs are characters that can't be manipulated directly by gamers, but are controlled by computers (game operating systems). For example, an agent NPC or the like that is manipulated by a computer by a counterpart of a gamer manipulated by a computer or a gamer can be specified on behalf of the computer.

In the human world, gamers manipulate the game through keyboard or mouse manipulation in certain game states. Each user has their own unique game style. This game style is a factor influencing the game operation of the gamers. In the computer world, gamers' game styles are interpreted as human genetic information, divided into detailed categories of styles, and given values to express them as style vectors (S).

In addition, the process of operating a game in the human world is a change of game state over time. In computer world, such a game state is defined as a weight vector (W) by defining the state component and the number of divisions that separate each state component.

In addition, in the human world, gamers input commands in a specific game state according to their game styles, and appear as game outputs in the computer world, and express each game output item as an output vector (O).

FIG. 2 shows an example of the configuration of the style vector S, the weight vector W representing the game state, and the output vector O transformed in the computer world shown in FIG.

The style vector S represents gamers' game habits or propensities, for example, aggression, defensiveness, input frequency, stability, and the like. Style vector (S) has a feature that can be found more detailed and intelligent game style gamers as the number of items that distinguish the user's game style.

The weight vector (W) defines a game state, and each game state is My Action (Action1), Opponent's Action (Action2), My Health (Power1), Opponent's Health (Power2), and Me and the opponent's distance (Distance) It is defined as a component of 5-dimensional item composed of. In addition, it is assumed that each component has a division number of 15, 15, 3, 3, and 10 levels. In the Dalian game played by gamers, the game state is determined by changing the above five game components according to the passage of time or the input operation of the gamers. When each component is divided by the number of divisions, my action (Action1) is divided into 15 ranges, and the remaining components are also divided by the number of divisions. Are distinguished. Therefore, the total game states generated by dividing the five-dimensional components, respectively, are 15 * 15 * 3 * 3 * 10 = 20,250, which has a total of 20,250 game states.

The output vector O is a set of game outputs that cause a transition from a particular game state to the next game state. This occurs by the player entering the best command through the input device among the possible input operations in a particular game state.

In the game style transplant system 1 according to the present invention, the progress of the game is defined as the relationship between the style vector S, the weight vector W, and the output vector O. Since the model is based on the empirical information that the actual gamers play the game, the complexity is greatly lowered compared to the above-described prior art (especially 2003-72640). This is because the above-described prior art forms a blind search space based on the input frequency and probability of the pure gamers.

However, in the present invention, the complexity is reduced through a predetermined algorithm in the limited definition region consisting of the style vector S, the weight vector W, and the output vector O, thereby reducing the load on the system. However, since the number of items (m) of the output vector (O) is determined by the game in advance, the more intelligent and actual gamers are, the more the number of items (n) of the style vector (S) and the number of state spaces (c) of the weight vector (W) are subdivided. It can be close to the game method of, but this becomes a factor that forms an explosive increase in complexity as described above.

3 shows the functional relationship between the style vector S, the weight vector W and the output vector O, which are defined through FIGS. 1 and 2 and used in the technical construction according to the invention.

The style vector S is information obtained by dividing and quantifying the game progression of gamers into detailed items. It is like a set of unique user styles that determine the output of a game as if it were human genetic information. The style vector S is mapped to a specific output vector O through a calculation with a weight vector W in a specific game state. Expressed in terms of human world, the gamer selects a certain operation from among game operations dependent on a specific game state (VS.weight vector) according to his game style (VS.style vector). It is to generate branch game output (VS. output vector).

A total of 20,250 game states consist of a two-dimensional weight vector (W) matrix that maps the style vector (S) to an arbitrary output vector (O). The k column in the weight vector W causes a specific output vector Ok. That is, the weight vector is information defining a relationship (correlation) between the style vector S and the output vector O in a specific game state.

The output vector O is a set of possible outputs obtained through the calculation of the style vector S and the weight vector W of a specific game state. The computed value of that possible output is related to the probability that the output will occur in the actual game. Therefore, the output vector value of the possible output having the highest value can be defined as the highest occurrence probability. This functional relationship is expressed as an expression as follows.

Style Vector (S): Vector S (n)

Output Vector (O): Vector O (m)

Weight vector (W): Vector W (n * m)

1 <= i <= n, 1 <= k <= m, 1 <= j <= m,

According to the defined n, m, the weight vector W exists in the two-dimensional matrix of n * m as many as the total number of state spaces (20,250). Therefore, the j column of the weight vector W stores the association information for determining the specific output vector Oj with respect to the style vector S.

4 illustrates the concept of a learning and prediction environment through the relationship between each vector in the technical scheme according to the present invention.

First, the game style transplant system 1 learns the weight vector W based on game log information performed by test gamers. In order to learn the weight vector W, a tester's personal style vector S must be set in advance. After the learning of the weight vector W, the weight vector W has generalized association information for generating the output vector O from the style vector S of various test gamers. That is, the game progress method according to the game log of the test gamers is generalized for each style and built in the game style transplant system 1 (learning weight vector from the test gamers).

Next, the game log information of the game object to be transplanted is analyzed and the learning of the style vector S is performed inversely from the output vector O corresponding to the game output included in the game log information. Style vector learning).

Finally, the style vector (S) reaches the convergence value and finishes learning. The output vector (O) is calculated by calculating the style vector (S) of the trained target gamer and the weight vector (W) learned by generalization. Create When the game output corresponding to the highest value in the output vector is delivered to the NPC, the NPC performs a game in which the game style of the transplanted gamers is reflected (style transplantation of the transplanted gamers to the NPC).

5 illustrates a configuration of a game style transplant system according to an embodiment of the present invention. The game style transplant system 1 includes a vector definition module 10, a game log generation module 20, a style vector digitization module 30, a weight vector learning module 40, an individual gamer style vector learning module 50, and a game. And an output inference module 60 and a game adapter module 70.

The vector definition module 10 defines the style vector S, the weight vector W, the number of items of the output vector O, the matrix information, and the like, and stores them in the vector definition DB 11. The style vector (S) is information that quantifies the psychological information, habits, and abilities when the game is played. The weight vector W is information obtained by quantifying an association relationship in which a specific output vector O is caused by the gamer's style vector S for each particular game state. And, the output vector (O) is information that digitizes the game output that can be output by the gamer through the operation of the input device in a specific game situation as a game action.

The game style transplantation system 1 according to the present invention is expandable to be applicable to various games without being dependent on a specific game. Therefore, the vector definition module 10 may set the number of items of each vector to fit a specific game and apply it to a specific game. For example, in a competitive action game, role playing game (RPG), and strategy simulation game, the user's game style is itemized to define a style vector (S), and the game state is totally determined through the item components and the number of divisions. By defining a game state space and defining an output vector O that can be output in a specific game state, a weight vector W of the entire game state can be defined. Due to such scalability, the game style transplantation system 1 according to the present invention has an advantage that it can be applied to various games independently regardless of the type of a specific game.

The implementation of the technique according to the invention is divided into three stages. In the first step, the game style transplantation system 1 is constructed by setting the style vector S of the test gamers and learning the weight vector W (learning the weight vector from the test gamers). The second step is to learn the user's style vector S from the game output O of the player to be transplanted through the game style transplant system 1 already established (style vector learning from the player to be transplanted). The third step is to infer (predict) the game output by calculating the style vector (S) in the second step and the weight vector (W) in the first step, and to transplant the style of a specific gamer to the NPC (object to be transplanted to the NPC). Porter's style).

The structure of step 1 is to play the game similar to the process of the game log by learning the weight vector (W) so that the tester's style vector (S) can be related to the output vector (O) through the game log of the test gamers. It is a structure for constructing a style porting system (1).

The game log generation module 20 receives game information of test gamers executed in the game operating system 100 through the game adapter module 70 and stores the game information in the game log DB 21.

The style vector numerical module 30 digitizes the individual style vector S of the game log DB 21 and stores it in the style vector DB 31. One of the following two modules is optionally performed in the digitization process. The survey digitization module uses a traditional technique of surveying test gamers and numerically clustering the survey results through the SOFM technique into a style vector (S). The other is to quantify the style vector S for each game style item by analyzing the game log stored in the game log DB 21 as a game log digitizing module.

The weight vector learning module S40 learns the weight vector W through repetitive learning between the style vector S and the output vector O of the individualized tester.

The two-stage structure obtains the output vector (O) from the game output (O) generated from the game log of the transplant target gamemer, and conversely, the transplant target gamer through the weight vector (W) learned in step 1 from the output vector (O). It is to learn the style vector (S) of.

The individual gamer style vector learning module 50 is based on the output vector (O) that matches the game output of the game player to be transplanted obtained from the game log DB 31 and the current style vector (S) and weight vector (W). Iteratively learns the style vector (S) of a specific gamer through the operation. According to the iterative learning, the value of the style vector reaches a convergence value.

In addition, the individual gamer style vector learning module 50 analyzes the game log of the gamer to be transplanted and digitizes the style vector for each game style item by the game log digitizing module of the style vector digitizing module 30 together with the above-described calculation function. It is also possible. Both can be selected by the user to determine the learning style.

The three-stage structure is to transfer the game output to the NPC of the game operating system and to port the game behavior according to the style of the target gamer to the NPC.

The game output inference module 60 infers the output vector (O) through a matrix operation between the learned style vector (S) of the second stage and the weight vector (W) of the learned stage, and matches the output vector (O). The game output is ported to the NPC of the game operating system 100 through the game adapter module 70.

Hereinafter, the learning in the game style transplant system 1 according to the present invention will be described in detail.

In particular, in the learning of the game style transplant system 1 according to the present invention, a newly learned vector value is obtained through the learning coefficient A obtained by an exponentially decreasing function to influence the speed and the result of learning. . The following is the equation of the learning coefficient (A).

A = a + ((A-a) * d)

A: learning coefficient,

a: minimum value

d: reduction rate

The above function has the characteristic of decreasing exponentially starting from the value of setting the initial learning coefficient (A), and it is easy to adjust the reduction rate by adjusting the reduction rate (d). In addition, since Equation 2 is an empirical formula based on empirical information, more efficient learning can be performed by adjusting learning environment variables (A, a, d) constituting the empirical formula based on the experience obtained through the learning process.

6 shows a detailed configuration of the weight vector learning module 40 according to an embodiment of the present invention. The weight vector learning module 40 includes a game state obtaining module 42, a game output obtaining module 44, a style vector obtaining module 46, a weight calculation processing module 47, and a plurality of DBs referenced by these modules. It is configured to include.

The game state obtaining module 42 obtains a game state identification code corresponding to the game state of the test gamer read from the game log DB 21 from the game state definition DB 43. The game state definition DB 43 stores a game state identification code that matches the game state defined by the number of divisions for each state component.

The game output acquisition module 44 obtains a game output identification code corresponding to the game output of the test gamer read from the game log DB 21 from the game output definition DB 45. Game output definition DB 45 stores a game output identification code that matches the game output possible by the user's input operation in a particular game state.

The style vector acquisition module 46 acquires the style vector S of the individual test gamer from the style vector DB 31 constructed by the style vector digitization module 30.

The weight calculation module 47 obtains the current weight vector W (t) corresponding to the game state obtained by the game state obtaining module 42 from the weight vector DB 41, and the current weight vector W (t). ) And the style vector S obtained by the style vector acquisition module 46 are generated as a new weight vector W (t + 1) through operation, and stored in the weight vector DB 41 to repeat the learning.

In particular, the weight calculation module learns the weight vector W such that the actual game output O extracted from the game log matches the output vector Oj having the maximum value after the calculation, and the equation is as follows.

t: current time

t + 1: Newly learned point from the current point

A: Learning coefficient

7 illustrates a configuration of an individual gamer style vector learning module 50 according to an embodiment of the present invention. The individual gamer style vector learning module 50 may include a game state acquisition module 42, a game output acquisition module 44, a style vector acquisition module 46, a style calculation processing module 51, and a weight vector DB 41. It is configured to include. The description common to the weight learning module 40 in FIG. 6 is omitted.

The game state obtaining module 42 obtains a game state identification code corresponding to the game state of the test gamer read from the game log DB 21 from the game state definition DB 43.

The game output obtaining module 44 obtains a game output identification code corresponding to the game output of the test gamer read from the game log DB 21 from the game output definition DB.

The style vector acquisition module 46 acquires the style vector S of the individual test gamer from the style vector DB 31 constructed by the style vector digitization module 30.

The style calculation processing module 51 obtains the current weight vector W corresponding to the game state obtained by the game state obtaining module 42 from the weight vector DB 41, and the style and the current weight vector W (t). The vector acquiring module 46 generates the style vector S obtained through the calculation as a new weight vector W (T + 1) and stores the weight vector DB 41 again in the weight vector DB 41 to repeat the learning.

The style operation processing module 51 obtains the weight vector W corresponding to the game state obtained by the game state obtaining module 42 from the weight vector DB 41, and the weight vector W and the style vector obtaining module 46. ) Generates a new style vector (S (t + 1)) through the operation of the current style vector (S (t)) of the player to be transplanted, and stores the result again in the style vector DB (31) to repeat the learning. .

In particular, the style calculation processing module 51 learns the style vector S so that the actual game output O extracted from the game log matches the output vector Oj having the maximum value after the calculation, and the equation is as follows.

8 is a flowchart illustrating a game style transplantation method according to an embodiment of the present invention. As in the game style transplant system 1 described above, the transplant method is divided into three stages. Step 1 is a weight vector learning process of steps S20 to S40, step 2 is a style vector learning process of step S50, and step 3 is a style transplantation process of step S60.

In the vector definition and storage step S10, the items of the style vector S, the weight vector W for each game state, and the output vector O are defined and stored in the vector definition DB 11. This step helps to port the user's style in the general game by defining the number of style items, game states, and outputs according to the type of game.

Next, step S20 to step S40 are steps of constructing a game style transplant system 1 from game logs of test gamers. Through this process, a number of gamers model the game output according to their styles, and the weight vector (W) serving as a function between the style vector (S) and the output vector (O) is learned and holds generalized information. Done.

Game log generation step (S20) generates a game log by storing the information in the game log DB 21, the process of the plurality of test gamers repeatedly performed the game. In this process, the gamer plays an actual game, and game information including gamers 'ID, game time, game state information of the gamers, and game outputs according to the gamers' input operation is stored.

In the test gamer style vector setting step (S30), the game style of the individual test gamers is analyzed from the game log, and the style vector is digitized and stored in the style vector DB 31.

In the weight vector learning step (S40), the weight vector W that causes a specific output vector O in a specific game state is learned by the style vectors S of the plurality of test gamers. This trains the weight vector W such that the style vector S of the test gamers matches the output vector O obtained from the game log. The learned weight vector W has a generalized and formal mapping information as an association function for mapping between the style vector S and the output vector O of the user.

After completing the above steps, the game style porting system 1 is built, and the porting system can infer the game style of a specific gamer. In the style vector learning step (S50), the game player's style vector S is inversely learned according to the game output inputted by the gamer from the game log of the game player to be transplanted.

In particular, the style vector learning step (S50) is preferably implemented as a target for porting the current gamers and real-time learning the game style according to the input operation of the gamers while the game is progressing while generating a real-time log file.

Next, in the gamer style transplantation step (S60), the game state is received from the game operating system 100, and the output vector O is inferred and determined based on the learned style vector S of the target gamer. The game output corresponding to the output vector (O) is transmitted to the game operating system 100 to transplant the game style of a specific gamer to the NPC.

9 shows detailed steps of the vector definition and storage step S10 according to an embodiment of the present invention.

The style vector definition step S11 defines style items that define psychological information, habits, and abilities when the gamers progress the game as style vectors.

The weight vector definition step S12 defines a weight vector composed of weights representing an association relationship in which a specific output vector is caused by the gamer's style vector for each specific game state.

The output vector definition step S13 defines a game output item that can be output by the gamer as a game action through manipulation of the input device in a specific game situation as an output vector.

10 is a detailed flowchart of a test gamer style vector setting step S30 according to an embodiment of the present invention.

Survey quantization step (S31) analyzes the user style through the survey and the result is numerically clustered through the SOFM (Self-Organizing Feature Map) technique to digitize the style vector.

It is also possible to selectively quantize the style vector for each game style item by analyzing the game log stored in the game log DB 21 by selectively adopting the game log digitizing step (S32).

11 shows a detailed procedure of the weight vector learning step S40 according to an embodiment of the present invention.

In the learning environment variable initial value setting step (S41), the learning environment variables (A, a, d) affecting the speed and quality of learning are initialized. In the weight vector initial value setting step (S42), an arbitrary initial value is set in preparation for learning.

Next, in the game log information reading step (S43) to read the game log of the test gamers one by one. In the game state obtaining step (S44), the game state is obtained from the read game log. In the current weight vector reading step S45, the acquired weight vector W of the current game state is read. In the new weight vector generation and storage step S47, the weight vector W is repeatedly learned according to Equations 3 and 4 described above.

After one lesson, the game state changes according to the game log of the same test player as the next game log is read, and then the next game player is transferred. If the maximum number of repetitions or convergence is reached according to the repeated learning, the learning is terminated. Otherwise, the learning coefficient A is adjusted through Equation 2 in the learning coefficient adjustment step (S48). Reread log DB21.

12 is a detailed flowchart of a weight vector generation and storage step S47 according to an embodiment of the present invention.

In the output vector acquisition step S471, game output is extracted from game logs of test gamers. The output vector difference calculation step (S472) calculates the difference between the output vector (O) corresponding to the actual game output and the output vector inferred by the operation. In operation S473, the output vector difference and the style vector are multiplied by the style vector. Equation 3 is used in the process between steps S471 to S473.

13 is a detailed flowchart of a style vector learning step S50 according to an embodiment of the present invention.

In the learning environment variable initial value setting step (S51), the learning environment variables (A, a, d) affecting the speed and quality of learning are initialized. In the style vector initial value setting step S52, an arbitrary initial value is set in preparation for learning.

Next, in the game log information reading step (S53) to read the game log of the transplant target gamers one by one. In the game state obtaining step (S54), the game state is obtained from the read game log. In the weight vector reading step S55, the weight vector W of the already learned game state is read. In the new style vector generation and storage step S57, the style vector S is repeatedly learned according to Equations 5 and 6 described above.

If the maximum number of repetitions or convergence is reached according to the repeated learning, the learning is terminated. Otherwise, the learning coefficient A is adjusted through Equation 2 in the learning coefficient adjustment step (S48). Reread log DB21.

14 is a detailed flowchart of a new style vector generation and storage step S57 according to an embodiment of the present invention. In the output vector acquisition step S571, game output is extracted from game logs of gamers to be transplanted. The output vector difference calculation step S572 calculates a difference between the output vector that matches the actual game output and the output vector inferred by the calculation. In operation S573 of output vector difference and weight vector, the weight vector is multiplied by the difference obtained in step S572. Equation 4 is used in the process between steps S571 to S573.

In particular, in the game style learning step of the game player to be transplanted, a method of calculating the inverse from the game output of the game log is possible, and adopts the game log digitization step (S32) of the test gamer style vector setting step (S30) shown in FIG. By analyzing the game log stored in the game log DB (21) it is also possible to digitize the style vector for each game style item. Both are optional.

15 shows a detailed procedure of the gamers style implantation step (S60) according to an embodiment of the present invention.

In the game state receiving step (S61), the game state of the game player to be transplanted is received from the game operating system 1. In the style vector reading step (S62), the style vector S of the player to be transplanted is read. In the weight vector reading step S63, the weight vector W that has already been learned is read through the game log of the test gamers.

In the output vector generation step S64, an output vector O is generated from the style vector S and the weight vector W read using Equation 1. In the game output transmission step (S65), the game output corresponding to the highest value of the generated output vector (O) is transmitted to the game operating system (1).

Since the game operating system 1 operates the NPC based on the received game output, the NPC performs game behavior according to the game style of the gamer to be transplanted.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited thereto and is intended by those skilled in the art to which the present invention pertains. Of course, various modifications and variations are possible within the scope of equivalents of the claims to be described.

FIG. 16 is a graph illustrating a style graft analysis result by a gamer style graft system and method according to an embodiment of the present invention.

The behaviors of gamers and the results of NPC transplantation that are the subject of learning and transplantation are as follows.

A) Learning about the situation where gamers are also stopped when no event occurs in the game

=> NPC is standing still in the game progress.

B) Learning game behavior to move forward and backward in gamers as the game progresses

=> NPC also moves back and forth during the game.

C) Gamer learns to play a game of punching and kicking an opponent

=> NPCs also attack their opponents with fist / foot attacks.

D) Learning to defend only against opponent character's attack during the game

=> NPCs only defend themselves.

E) Learning movement, attack, and defense action according to the gamers' style during the game

=> NPC moves / attack / defends depending on the situation during the game

The graph shown in FIG. 16 shows the results of learning and transplanting by the A, B, C, D, and E scenarios above. The maximum number of repetitions was 100 as a condition of weight vector (W) learning. In addition, the game state was limited to My Action (Action1), the opponent's Action (Action2), and the distance between me and the opponent (Distance). In addition, the learning rate was expressed as a percentage of the match between the game output recorded in the game log of the transplanted gamer and the game output of the transplanted NPC.

The results showed that learning about scenario A showed a 100% match rate, and the match rate for each scenario was above 85% on average.

17A and 17B show examples of components constituting a game state when applying the technique of the present invention to other kinds of games. FIG. 17A shows an example of a state splitting element when targeting a role playing game (RPG), and FIG. 17B shows an example of a state splitting element when targeting a strategy simulation game.

The user defines the game state by defining the components and the number of divisions constituting the game state according to the type of the game, and defines the number of items analyzing the user's style to apply the technology of the present invention to various kinds of games. It is possible.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited thereto and is intended by those skilled in the art to which the present invention pertains. Of course, various modifications and variations are possible within the equivalent law of the claims to be described.

According to an aspect of the present invention, by analyzing the game style of the gamers with a game style vector and a weight vector, it is possible to build a game system that reduces the load on the process of the game system through a limited vector operation in a limited definition region.

According to another aspect of the present invention, by providing a game system in which the artificial intelligence that implants the game style of a particular gamer is active, it is possible to play the game on behalf of the artificial intelligence in the absence of the gamer.

According to another aspect of the present invention, it is possible to provide a more dynamic game service by making it possible to compete with the general gamers of artificial intelligence implanted with game styles of famous pro gamers.

The following drawings attached to this specification are illustrative of preferred embodiments of the present invention, and together with the detailed description of the invention to serve to further understand the technical spirit of the present invention, the present invention is a matter described in such drawings It should not be construed as limited to

1 is a conceptual diagram of a learning model based on the gamers constituting the technology of the present invention.

2 is an exemplary diagram of a style vector, a weight vector, and an output vector model constituting the technique of the present invention.

3 is a relationship diagram between a style vector, a weight vector, and an output vector constituting the technique of the present invention;

4 illustrates the concept of a learning and transplant environment through the relationships between the respective vectors making up the technique of the present invention.

5 is a block diagram of a game style transplant system according to an embodiment of the present invention.

6 is a block diagram of a weight vector learning engine according to an embodiment of the present invention.

7 is a block diagram of a style vector learning engine according to an embodiment of the present invention.

8 is a flow chart of a game style transplant method according to an embodiment of the present invention.

9 is a flow chart of the vector definition and storage step according to an embodiment of the present invention.

10 is a flowchart illustrating a test gamer style vector setting step according to an embodiment of the present invention.

11 is a flowchart of a weight vector learning step according to an embodiment of the present invention.

12 is a flow chart of the weight vector generation and storage step according to an embodiment of the present invention.

Figure 13 is a flow chart of the style vector learning step according to an embodiment of the present invention.

14 is a flow chart of a new style vector generation and storage step according to one embodiment of the invention.

15 is a flow chart of a gamer style implantation step in accordance with one embodiment of the present invention.

16 is a graph of results analysis of style graft according to one embodiment of the present invention.

17A and 17B are exemplary diagrams of game state components when applied to different types of games (Role Playing Game, Real Time Simulation Game).

Claims (21)

In the game style transplant system that learns the game style of the target gamers transplanted to the NPC, A vector definition module for storing items of a style vector, a weight vector for each game state, and an output vector in a vector definition DB; A game log generation module for generating a game log by informing a process of performing a game by a plurality of test gamers and storing the information in a game log DB; A style vector digitizing module for analyzing and digitizing game styles of individual test gamers stored in the game log DB and storing them in a style vector DB; A weight vector learning module for learning the weight vector for each general game state by storing the weight vector learned from the current weight vector in the weight vector DB based on the game output of the test gamers stored in the game log DB; An individual gamer style vector learning module for learning the inferred style vector by storing the style vector derived based on the game log of the game player to be transplanted into the style vector DB; A game output inference module for predicting and determining an output vector from a style vector learned by the individual gamer style vector learning module and a weight vector of a current game state for a player to be transplanted; And It is dependent on a specific game operating system, receives the current game state from the game operating system, and matches the output vector generated by the game output inference module in order to port the game style of the specific gamer to the NPC of the game operating system. Game adapter module for delivering game output to the game operating system Game style transplant system of gamers through artificial intelligence learning, comprising a. The method of claim 1, The vector definition module, Style items that define psychological information, habits, and abilities when gamers play the game are defined as style vectors, For each particular game state, define a weight vector consisting of weights that represent the associations caused by the output vector by the player's style vector, By defining game output items as output vectors that gamers can output as game actions through manipulation of input devices in certain game situations, Game style porting system of gamers through artificial intelligence learning, which can be applied to various games independently of a specific game type. The method according to claim 1 or 2, The game state is, The total number of game states is determined according to the state components defining the progress of the game and the number of divisions given per unit of the state components. Game style transplantation system of the gamers through the artificial intelligence learning, characterized in that the more the number of the state component and the number of partitions to form a detailed game state. The method of claim 1, The style vector digitization module, A survey digitizing module for surveying test gamers and digitizing the survey results through a SOFM technique into a style vector; And Game log digitization module for digitizing the style vector for each game style item by analyzing the game log stored in the game log DB Game style transplantation system of the gamers through artificial intelligence learning, characterized in that by selecting any one of the test data stored in each of the gamer digitized in the style vector DB. The method of claim 1, The weight vector learning module, A game state definition DB containing a game state identification code matching the game state defined by the number of divisions by state component; A game state acquisition module for obtaining a game state identification code corresponding to a game state of a test gamer read from the game log DB from the game state definition DB; A game output definition DB that contains a game output identification code that matches a game output by a user's input operation in a specific game state; A game output acquisition module for obtaining a game output identification code corresponding to a game output of a test gamer read from the game log DB from the game output definition DB; A style vector DB for storing a style vector of the individual test gamers analyzed by the style vector digitizing module; A style vector acquisition module for obtaining a style vector from the style vector DB; A weight vector DB for storing a weight vector for each game state for mapping the style vector to an output vector for each specific game state; And Obtaining a current weight vector corresponding to the game state obtained by the game state obtaining module from the weight vector DB, and generating the new weight vector through operation by calculating the current weight vector and the style vector acquired by the style vector obtaining module; Weight calculation processing module for repeating the learning by storing in the weight vector DB again It is made, including The game style transplantation system of the gamers through artificial intelligence learning, characterized in that the weight vector is repeatedly learned through the calculation of the weight calculation module based on the game information of the test gamers read from the game log DB. The method of claim 5, The weight calculation processing module, A style vector obtained by the style vector obtaining module; A current weight vector obtained from the weight vector DB with respect to a game state obtained by the game state obtaining module; And Output vector obtained from the game output according to the game state of the test gamer read from the game log DB Game style transplantation system of gamers through AI learning, characterized in that generating a new weight vector from. The method of claim 1, The individual gamer style vector learning module, A game state definition DB containing a game state identification code matching the game state defined by the number of divisions by state component; A game state acquisition module for acquiring a game state identification code corresponding to a game state of a game player to be transplanted read from the game log DB from the game state definition DB; A game output definition DB that contains a game output identification code that matches a game output by a user's input operation in a specific game state; A game output acquisition module for obtaining a game output identification code corresponding to a game output of a test gamer read from the game log DB from the game output definition DB; A style vector DB for storing a style vector of the game player to be transplanted; A style vector acquisition module for obtaining a current style vector from the style vector DB; A weight vector DB for storing a weight vector for each game state for mapping the style vector to an output vector for each specific game state; And Obtain a weight vector corresponding to the game state obtained by the game state obtaining module from the weight vector DB, and calculate a new style vector by calculating the weight vector and the current style vector of the transplanted gamers acquired by the style vector obtaining module. Style calculation module for repeating the learning by generating and storing in the style vector DB again It is made, including Based on the game information of the game player to be transplanted read from the game log DB, the style vector is repeatedly performed through the calculation of the style operation processing module, and the style vector value of the game player to be transplanted is reversed from the patterned weight vector value. Game style transplantation system of gamers through artificial intelligence learning, characterized in that learning by converging. The method of claim 1, The individual gamer style vector learning module, Game style porting system of the gamers through the artificial intelligence learning, characterized in that by analyzing the game log of the game object to be transplanted by the game log number module included in the style vector numerical module to quantize the style vector for each game style item. The method according to claim 7 or 8, The individual gamer style vector learning module, Game style transplantation system of the gamers through artificial intelligence learning, characterized in that to generate a game log progressed by the gamers of the transplant target in real time to learn the game style vector of the transplant target gamers in real time. The method of claim 7, wherein The style operation processing module, A style vector obtained by the style vector obtaining module; A weight vector learned from test gamers by the weight vector learning module; And Output vector obtained from the game output according to the game state of the test gamer read from the game log DB Game style transplantation system of gamers through AI learning, characterized in that generating a new weight vector from. The method of claim 6 or 10, A game style transplantation system for gamers through artificial intelligence learning, which obtains a newly learned vector value through a learning coefficient obtained by an exponentially decreasing function to influence the speed and result of learning. In the game style porting method of learning the game style of the target gamers transplanted to the NPC, (1) defining items of a style vector, a weight vector for each game state, and an output vector and storing the items in a vector definition DB; (2) generating a game log by storing information in a game log DB of a plurality of test gamers repeatedly performing a game; (3) analyzing and quantifying the game styles of the individual test gamers stored in the game log DB and storing them in the style vector DB; (4) learning a weight vector causing a specific output vector in a specific game state by style vectors of a plurality of test gamers; (5) learning a style vector of the gamer according to a game output inputted by the gamer from the game log of the game player to be transplanted; And (6) Receive the game state from the game operating system, infer the output vector based on the learned style vector of the gamer to be transplanted, determine the output vector, and deliver the output value corresponding to the output vector to the game operating system game of the specific gamers Steps to port styles to NPC Game style transplantation method of the gamers through the AI learning, characterized in that comprises a. The method of claim 12, Step (1), (1-1) defining style items that define psychological information, habits, and abilities as gamers play the game as style vectors; (1-2) defining a weight vector composed of weights representing association relations in which a specific output vector is caused by a gamer's style vector for each specific game state; And (1-3) The step of defining a game output item as an output vector that the gamer can output as a game action through manipulation of the input device in a specific game situation. Consisting of, Game style porting method of gamers through AI learning, characterized in that it can be applied to a variety of games independently of a specific game type. The method of claim 12, Step (2) is, Game style porting method of gamers through artificial intelligence learning, characterized in that the game log including the game user ID, game time, game state information of the gamers, game output information according to the input operation of the gamers in the game log DB . The method of claim 12, Step (3), (3-1) surveying the test gamers, and clustering the survey results through a SOFM technique to digitize the styles into style vectors; And (3-2) analyzing the game log stored in the game log DB to digitize the style vector for each game style item; The game style transplantation method of the gamers through artificial intelligence learning, characterized in that by selecting any one of the steps to store the quantized style vector for each test gamer in the style vector DB. The method of claim 12, The step (4), (4-1) setting an initial value to a learning environment variable; (4-2) setting an arbitrary initial value in the weight vector DB; (4-3) reading game log information of a test gamer from the game log DB; (4-4) obtaining a game state from the game log information; (4-5) reading the current weight vector according to the game state from the weight vector DB; (4-6) reading the style vector of the gamer from the gamer style DB constructed in the step (3); And (4-7) generating a new weight vector through operation from the current weight vector of the step (4-5) and the style vector of the step (4-6) and storing the new weight vector again in the weight vector DB for learning Consisting of, Game style porting method of the gamers through the artificial intelligence learning, characterized in that for learning the weight vector repeatedly through the operation based on the game information of the test gamers read from the game log DB. The method of claim 16, Step (4-7), (4-7-1) obtaining an output vector corresponding to the game output of the gamer from the game log DB; (4-7-2) obtaining a difference between the output vector of the step 4-7-1 and the output vector currently inferred; (4-7-3) A step of obtaining a new learning value by multiplying the difference vector obtained in the step (4-7-2) by the gamer's style vector Game style transplantation method of the gamers through the AI learning, characterized in that comprises a. The method of claim 12, The step (5), (5-1) setting an initial value to a learning environment variable; (5-2) setting the game style vector of the gamer to be transplanted to the style vector DB as an initial value; (5-3) reading game log information of a test gamer from the game log DB; (5-4) obtaining a game state from the game log information; (5-5) reading the weight vector according to the game state from the weight vector DB; (5-6) reading the current style vector of the gamer from the gamer style DB constructed in the step (5-2); And (5-7) generating a new style vector through operation from the weight vector of step (5-5) and the current style vector of step (5-6), and storing the new style vector again in the style vector DB for learning Consisting of, Game style transplantation method of the gamers through artificial intelligence learning, it characterized in that the learning of the style vector repeatedly through the operation based on the game information of the target gamers to be read from the game log DB. The method of claim 18, Step (5-7), (5-7-1) obtaining an output vector corresponding to the game output of the gamer from the game log DB; (5-7-2) obtaining a difference between the output vector of the step 5-7-1 and the output vector currently inferred; (5-7-3) A step of obtaining a new learning value by multiplying the generalized weight vector by the difference obtained in the step (5-7-2) Game style transplantation method of the gamers through the artificial intelligence learning, characterized in that it comprises a repetitive learning the style vector. The method of claim 16 or 18, The learning environment variable, The learning coefficient is obtained by a function that exponentially decreases to affect the speed and the result of learning, and the learning coefficient is adjusted by re-learning after each learning is performed through all the logs of the game log DB. Game style porting method of gamers through artificial intelligence learning, characterized in that to make. The method of claim 12, The step (6), (6-1) receiving a game state of the NPC in the game operating system; (6-2) reading the style vector of the player to be transplanted learned in the step (5) from the style vector DB; (6-3) reading the weight vector according to the game state of step (6-1) from the weight vector DB learned in step (4); (6-4) generating an output vector by computing the style vector of step 6-2 and the weight vector of step 6-3; And (6-5) transmitting a game output corresponding to a vector value corresponding to the highest value among the output vectors to the game operating system It comprises a, and predicts the output through the calculation of the weight vector learned from the game log of the test gamers and the style vector learning the style of the target gamers to transfer to the game operating system characterized in that the transplant to the NPC Game style porting method of gamers through artificial intelligence learning.