CN115840840A

CN115840840A - Situation awareness plan optimal selection ordering method for learning focus

Info

Publication number: CN115840840A
Application number: CN202310136364.1A
Authority: CN
Inventors: 金欣; 王新年; 朱燕; 陈洪辉; 蔡飞
Original assignee: CETC 28 Research Institute
Current assignee: CETC 28 Research Institute
Priority date: 2023-02-20
Filing date: 2023-02-20
Publication date: 2023-03-24

Abstract

The invention discloses a method for optimizing and sequencing learning-concerned situation-aware plans, which can automatically screen and generate a reference plan recommendation list from a historical plan library according to emergency situations and key point information in a subject task indication, comprehensively consider factors in aspects of situation similarity, resource availability and effect superiority and inferiority 3 to carry out optimization sequencing, present interpretable comparative analysis and assist a director to quickly find a proper reference plan so as to accelerate plan formulation. Meanwhile, the method collects the operation of the user on the recommendation list, implicitly marks sample data, selects the focus of attention of the reference plan under various emergency situations of incremental learning, optimizes the focus of attention model on line, can gradually reduce the searching time of the optimal reference plan along with the use of the user, and greatly improves the working efficiency.

Description

Situation awareness plan optimal selection ordering method for learning focus

Technical Field

The invention relates to a decision support technology, in particular to a method for optimally ordering a situation awareness plan for learning focus.

Background

In application scenes such as military operations, emergency rescue and relief, security protection of major activities and the like, emergency treatment needs to be carried out on emergency situations, and the time is often very short. At present, plans are formulated for various hypothetical situations at ordinary times, and relevant plans are selected and adaptive adjustment is rapidly made according to current tasks and situations during use. With the daily accumulation, the size of the plan library is increasing, the time cost for finding the plan is also increasing, and the lower the reference value of the found plan is, the time spent on the adaptability adjustment is greatly increased. Therefore, the system is urgently needed to have the automatic plan recommendation function, and plans suitable for current tasks and situations can be quickly and accurately screened and recommended from the daily and monthly plan library within seconds.

In various emergency treatment command and control systems, a widely adopted plan recommendation method mainly comprises keyword search. The method is low in accuracy, and the search result still needs a user to spend a large amount of time for reading and screening, so that the method is not beneficial to rapid handling of emergency. Therefore, a system is needed to automatically extract the most suitable plan from the plan library as a reference plan according to the current emergency situation and the superior task instruction, generate a plan recommendation list at a second level, and have certain accuracy. The user can intuitively know all aspects of characteristics and corresponding evaluation of the recommended plan and quickly make a plan selection.

Disclosure of Invention

The purpose of the invention is as follows: the invention aims to provide a method for optimally ordering a learning and focusing context-aware plan, so that a reference plan is quickly recommended and accurately ordered for a commander in an emergency, and plan making is accelerated.

The technical scheme is as follows: the invention relates to a method for optimally ordering a situation awareness plan for learning focus, which comprises the following steps:

(1) Manually constructing a plan library P, and constructing a plan feature vector PC for each plan;

(2) Initializing and constructing a case selection focus model W;

(3) Extracting situation characteristics from a superior task instruction T and an emergency situation S when a task is received in emergency, and constructing a current situation characteristic vector SC;

(4) Respectively calculating 3 indexes of situation similarity, resource availability and effect superiority and inferiority for each plan in a plan library P, and screening according to a threshold value to form a plan recommendation list Re;

(5) Calculating a comprehensive recommendation index for each plan in the plan recommendation list Re, and sorting from high to low to form a plan sorting list Ra;

(6) A plan arrangement list Ra is presented on a human-computer interaction interface, the analysis content can be explained by comparing with multiple plan multiple elements, and a user browses and finally selects a plan as a reference plan B;

(7) Implicitly marking newly-added option sample data SaN according to the characteristics of the reference plan B, and adding the option sample data SaH into an option sample database SaH;

(8) And iteratively calling a preferred case attention key model W based on the sample data SaN of the newly added selected case.

The step (1) is specifically as follows:

(1.1) combing the common emergency treatment task mode M in the business field, aiming at each task mode M _k Artificially defining plan feature vector model MM _k The method comprises 3 characteristic variables, specifically:

(1.1.1) the first set of feature variables is the task context feature variable MMS _k The situation characteristics reflecting the change under the task style comprise the change characteristics of the emergency situation S and the change characteristics of the superior task instruction T;

(1.1.2) the second set of characteristic variables is resource usage characteristic variables MMC _k Reflecting the resource types frequently used in the plans under the task styles;

(1.1.3) the third set of characteristic variables is the Performance evaluation characteristic variable MME _k Reflecting the common index of the execution efficiency of the evaluation plan under the task style;

(1.1.4) MMS _k 、MMC _k 、MME _k Combining to form a task style M _k Is predicted by the planning feature vector model MM _k ；

(1.2) for each protocol P in the library P of protocols _i According to their task style M _k According to the corresponding plan feature vector model MM _k Respectively MMS therein _k 、MMC _k 、MME _k Carrying out assignment and constructing the plan P _i Is predetermined characteristic vector PM _i The method specifically comprises the following steps:

(1.2.1) for each MMS _kj Respectively carrying out assignment according to a superior task instruction S and an emergency situation T to form a PMS _i ；

(1.2.2) for each MMC _kj According to the predetermined plan P respectively _i Assigning data of specific type used by each resource to form PMC _i ；

(1.2.3) for each MME _kj According to the predetermined plan P respectively _i Evaluating the data of each index to form PME _i ；

(1.2.4) mixing PMS _i 、PMC _i 、PME _i Combining to form a feature vector model PM of the current plan P _i 。

The step (2) is specifically as follows:

(2.1) for each type of task style M _k Construction of case selection focus model W _k The method specifically comprises the following steps:

(2.1.1) for each MMS _kj 、MMC _kj 、MME _kj Defining the weight vector WS separately _kj 、WC _kj 、WE _kj Manually assigning values to them by experience, or averaging the values to ensure

；

(2.1.2) is MMS _k 、MMC _k 、MME _k Three vectors define a three-dimensional weight vector WA _k Manually assigning values to them, or averaging them, to ensure WA _k1 + WA _k2 +WA _k3 =1；

(2.1.3) reacting WS _k 、WC _k 、WE _k 、WA _k Combining to form a task style M _k Case selection and attention focus feature model W _k ；

(2.2) all task styles M _k Corresponding case selection attention key feature model W _k Forming a set W.

The step (3) is specifically as follows:

calling out a corresponding plan feature vector model MM according to the type of the claimed task when emergent handling the task in case of the emergent handling of the claimed emergency _k And respectively carrying out assignment according to the superior task instruction S and the emergency situation T to form a current task situation feature vector SC.

The step (4) is specifically as follows:

(4.1) according to the current task style M _k The system automatically calls out the plans belonging to the task style from the plan library P to form a subset PM of the plan library P _k ；

(4.2) for PM _k Each plan PM of _ki And calculating the situation similarity of the preset case feature vector and the current task situation feature vector SC, specifically comprising the following steps:

(4.2.1) for each MMS _kj Calculate it and SC _j The value similarity between them forms the situation similarity vector SS _i In particular per MMS _kj The similarity calculation method is related to a specific application scene, and specific problems need to be specifically designed;

(4.2.2) calculation of the plan PM _ki Is a situation similarity value Sim _i For all characteristic variables MMS _kj The situation similarity value SS _ij Weighted summation:

；

(4.3) for PM _k Each plan PM of _ki And calculating the resource availability according to the preset characteristic vector, specifically:

(4.3.1) for each MMC _kj Searching whether available resources of corresponding types exist in an available resource library to form a resource availability vector FA with a Boolean value _ij 1 represents present, 0 represents absent;

(4.3.2) calculation of the plan PM _ki Resource availability value Ava of _i For all characteristic variables MMC _kj Resource availability value of FA _ij Weighted summation:

；

(4.4) for PM _k Each of the plans PM _ki And calculating the superiority and inferiority of the effect according to the characteristic vector of the plan, specifically comprising the following steps:

(4.4.1) for each MME _kj Adding PM to _ki And PM _k The other plans in (1) are compared transversely to calculate EF _ij The method specifically comprises the following steps:

(4.4.1.1) finding PM _k Middle MME _kj Maximum value of (2)

And a minimum value->

And PM _ki The value of (D) V;

(4.4.1.2) calculating MME _kj Is taken as the value EF _ij ：

；

(4.4.2) calculating the plan PM _ki The effect figure of (1) good or bad Eff _i For all characteristic variables MME _kj Is taken as the value EF _ij Weighted summation:

；

(4.5) for PM _k Each of the plans PM _ki Then mix it with Sim _i 、Ava _i 、Eff _i And respectively comparing with corresponding threshold values, and adding all the plans larger than the threshold values into a plan recommendation list Re.

The step (5) is specifically as follows:

(5.1) recommendation column for planEach of the schedules Re in Table Re _i Calculating a composite recommendation index

；

(5.2) according to Rank _i The value of the variable is from high to low, to Re _i And (5) sorting to form a plan sorting list Ra.

The step (6) is specifically as follows:

(6.1) in the plan arrangement list Ra on the human-computer interaction interface, when a user clicks one of the plans Ra _i In the case of the names of (2), ra is shown in a list _i All plan feature values MM of _ij And corresponding evaluation value SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i For analysis and reference of users;

(6.2) in the plan arrangement list Ra on the human-computer interaction interface, when a user clicks a 'join contrast column' beside a plurality of plans and clicks a 'plan contrast analysis' button, each plan Ra is contrastively displayed in a list form _i All plan feature values MM of _ij And corresponding evaluation value SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i For analysis and reference of users;

(6.3) in the plan arrangement list Ra on the human-computer interaction interface, when a user selects one plan Ra _i And when the button of 'selecting as reference plan' is clicked, the plan is recorded as a reference plan B.

The step (7) is specifically as follows:

(7.1) creating a sample data SaN containing the task style M _k Three data items of a reference plan evaluation vector X and a recommended plan evaluation matrix Y;

(7.2) setting the current task style M _k The task style SaNM marked as sample data SaN of the selected case _k ；

(7.3) for reference protocol B, SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i Combined to form a reference planEvaluating the vector X;

(7.4) forming a set RaM by N (the value of N can be customized) recommended plans which are arranged at the front in the plan arrangement list Ra and are not selected as reference plans, and aiming at each of the recommended plans RaM _i ，SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i Combined to form a predetermined RaM _i Is estimated as vector Y _i Combining the evaluation vectors of all RaMs to form a recommended plan evaluation matrix

。

The step (8) is specifically as follows:

(8.1) for Y _i Each-dimensional feature evaluation value Y in (1) _ij Calculating N numbers of Y _ij To form a mean vector

：

；

(8.2) calculating the variance of Y in each feature dimension

：

；

(8.3) definition

And &>

Two scaling factors are used for adjusting the learning speed, and the emphasis model W of the optimal solution is adjusted in an iterative mode:

；

a computer storage medium having stored thereon a computer program which, when executed by a processor, implements a method of learning context-aware protocol preference ranking for focus of interest as described above.

A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements a method for learning context-aware protocol preference ranking of focus of interest as described above when executing the computer program.

Has the advantages that: compared with the prior art, the invention has the following advantages:

1. according to the method, on the basis that the current emergency situation and the superior task instruction are understood to a certain degree, the similar plan is recommended accurately, and compared with a method for searching the keyword of the plan name, the accuracy can be effectively improved;

2. on the basis of comprehensively considering the factors of situation similarity, resource availability and effect superiority and inferiority 3, the method accurately recommends the plan, ensures that the recommended plan is suitable for the current situation, and can generate better effect;

3. the method and the system can learn the focus of attention of the selected reference plan under various emergency conditions, optimize the focus of attention model on line, and gradually reduce the time for finding the optimal reference plan along with the use of the user.

Drawings

FIG. 1 is a schematic diagram of the overall method of the present invention;

FIG. 2 is a flow diagram of a context similarity algorithm of the present invention;

FIG. 3 is a flow chart of a resource availability algorithm of the present invention;

FIG. 4 is a flow chart of the performance goodness algorithm of the present invention;

FIG. 5 is a flow chart of a protocol preference ranking algorithm of the present invention;

FIG. 6 is a flow chart of the focus learning algorithm of the present invention.

Detailed Description

The technical scheme of the invention is further explained by combining the attached drawings.

As shown in fig. 1, a method for learning a context-aware plan optimal ranking of an emphasis point includes the following steps:

(2) Initializing and constructing a case selection focus model W;

(6) A plan arrangement list Ra is presented on a human-computer interaction interface, analysis contents can be explained by comparing with a plurality of plan multi-elements, and a user browses and finally selects a plan as a reference plan B;

The step (1) is specifically as follows:

(1.1) combing common emergency treatment task styles M in the business field, aiming at each task style M _k Artificially defining plan feature vector model MM _k The method comprises 3 characteristic variables, specifically:

The step (2) is specifically as follows:

(2.1) for each type of task style M _k To build a case-selection focus model W _k The method specifically comprises the following steps:

；

(2.1.2) is MMS _k 、MMC _k 、MME _k Three vectors define a three-dimensional weight vector WA _k Manually endowing it with experienceValue, or average assigned value, ensuring WA _k1 + WA _k2 +WA _k3 =1；

(2.1.3) reacting WS _k 、WC _k 、WE _k 、WA _k Combining to form a task style M _k The case selection and attention focus feature model W _k ；

The step (3) is specifically as follows:

The step (4) is specifically as follows:

(4.2) for PM _k Each of the plans PM _ki And calculating the situation similarity of the preset case feature vector and the current task situation feature vector SC, specifically comprising the following steps:

(4.2.1) for each MMS _kj Calculate it and SC _j The value similarity between them forms the situation similarity vector SS _i In particular per MMS _kj The similarity calculation method is related to the specific application scene of the method, and specific problems need to be specifically designed;

；

(4.3) for PM _k Each of the plans PM _ki And calculating the resource availability according to the preset characteristic vector, specifically:

(4.3.1) for each MMC _kj Searching whether available resources of corresponding types exist in the available resource library to form a resource availability vector FA with a Boolean value _ij 1 represents present, 0 represents absent;

(4.3.2) calculation of Preset PM _ki Resource availability value Ava of _i For all characteristic variables MMC _kj Resource availability value of FA _ij Weighted summation:

；

(4.4.1) for each MME _kj Adding PM to _ki And PM _k The other plans in (1) are compared transversely and the EF is calculated _ij The method specifically comprises the following steps:

(4.4.1.1) finding PM _k Middle MME _kj Maximum value of

And a minimum value->

And PM _ki The value of (D) V;

(4.4.1.2) calculating MME _kj Is taken as the value EF _ij ：

；

(4.4.2) calculation of Preset PM _ki The effect figure of (1) good or bad Eff _i For all characteristic variables MME _kj Is taken as the value EF _ij Weighted summation:

；

(4.5) for PM _k Each of the plans PM _ki Then mix it with Sim _i 、Ava _i 、Eff _i Respectively comparing with corresponding threshold values, and adding all the predetermined plans larger than the threshold values into the predetermined plansA recommendation list Re.

The step (5) is specifically as follows:

(5.1) recommending each of the plans Re in the list Re for the plans _i Calculating a composite recommendation index

；

The step (6) is specifically as follows:

(6.2) in the plan arrangement list Ra on the human-computer interaction interface, when a user clicks a 'join contrast bar' beside a plurality of plans and clicks a 'plan contrast analysis' button, each plan Ra is contrastingly displayed in a list form _i All plan feature values MM of _ij And corresponding evaluation value SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i For analysis and reference of users;

(6.3) in the plan arrangement list Ra on the human-computer interaction interface, when a user selects one plan Ra _i And when the button of selecting as the reference plan is clicked, the plan is recorded as a reference plan B.

The step (7) is specifically as follows:

(7.3) reference Pre-predictionTable B, its SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i Combining to form a reference plan evaluation vector X;

。

The step (8) is specifically as follows:

(8.1) for Y _i Each-dimensional feature evaluation value Y in (1) _ij Calculating N number of Y _ij To form a mean vector

：

；

(8.2) calculating the variance of Y in each feature dimension

：/>

；

(8.3) definition

And &>

Two scaling factors are used for adjusting the learning speed, and the iterative tuning optimization scheme focuses on the key model W: />

；

Next, the specific implementation of the present invention will be described in terms of 4 aspects for some key links in the above steps, taking a target fire fighting task in a military combat application scenario as an example.

Example 1: method for constructing plan database with complete elements

In order to implement the plan recommendation, firstly, a plan library is constructed manually, and the characteristic vector of the plan is constructed and stored in the plan library for each planned plan.

Each protocol contained 3 aspects of data: task context, resource usage, effect assessment. The task situation is input, the resource usage is output, the effect evaluation is feedback, and the cause, the effect and the evaluation form a complete plan data.

The plan features are a plurality of feature attributes extracted from the plan data in order to support the plan screening sorting based on feature matching. Factors in 3 aspects are considered in the aspect of the pre-arranged screening and sorting emphasis:

(1) Context similarity: whether the task situation aimed by the plan is similar to the current task situation or not. Similar problems have similar solutions, and finding solutions of similar problems is usually the most effective method under the conditions of lacking knowledge and lacking data;

(2) Resource availability: whether the resource design used in the protocol is appropriate for the current situation. If more resources in the plan are not available in the current situation, then the plan is also not available;

(3) The effect is excellent and inferior: how well the plan itself is. The quality of the effect is also an important consideration when there are multiple candidate plans.

Therefore, the plan features also include task context, resource usage, and effect evaluation 3 aspects.

(1) Task context data and features

Connotation: the task requirements and the task situation aimed at when the plan is made are input basis for the plan making.

Data composition: the method mainly comprises the aspects of task requirements and task situations 2. The general planning operation flow takes the 2-step analysis work as input, and basically covers necessary information which needs to be mastered when a plan is made.

Characteristic extraction: the task context feature is an abstract abstraction of task requirements and task situation, and describes key elements in the task and situation in a summary way through a few short attributes. The purpose of the refinement is mainly used for calculating the context similarity.

Different types of tasks, task requirements and task situations are completely different, and a task situation characteristic model needs to be designed according to each type of tasks in a customized mode.

The explanation is given by taking a target fire striking task in a military combat application scene as an example:

task requirement characteristics: basic characteristics such as target hitting, damage requirement, completion time limit and the like;

task situation characteristics: including current position of the target, speed and course, health status, intent of action, day and night distinction, meteorological conditions, etc.

(2) Resource usage data and features

Connotation: and according to the task context, conspiring the design result of using the resources for the action of the my party.

And (3) data composition: taking a target fire fighting task in a military combat application scenario as an example, which troops to go from, the type and number of platforms to go out, weapon ammunition/electronic devices to mount, and the like, for each action.

Characteristic extraction: the goal is not for similar protocol matching, as the current action protocol is not yet available. The method is mainly used for judging whether the content of the plan is suitable for the current situation. If the resources used in the plan are not available in the current vicinity at all, or if the resources are not available in close proximity, the plan itself is not suitable for the current situation.

(3) Effect assessment data and features

Connotation: the recommended protocol is only similar but not enough, and should be successful and effective. There are several ways to evaluate a plan for good or bad:

and (3) actually executing the effect: if the plan is actually executed, the execution result is the best evaluation basis;

simulation deduction result: there may be a portion of the plan deduced, if any, for reference;

theoretical calculation results are as follows: roughly estimating the effect of a plan according to an operational research method, and counting the time and resource consumption;

the plan validity is as follows: whether the plan passes the upper level approval or not;

author authority: how well the protocol maker is;

and (4) pre-arranged comment: if it is possible to establish a review area, different users are allowed to review the plan like commodity recommendation, and the advantages and the disadvantages of the plan are indicated, so that the method also has a certain reference value.

And (3) data composition: taking a target fire striking task in a military combat application scene as an example, the effect evaluation content comprises the following steps:

(1) the hitting effect is as follows: hit probability to target using missile

(2) Time consumption: overall time consuming of all combat actions

(3) Resource consumption: total number of missiles launched for all operational actions

Characteristic extraction: the purpose of extracting features is mainly to calculate the superiority and inferiority of the effect and to use the superiority and inferiority in the plan sorting. The 3 data can be directly extracted for calculation.

Example 2: similar plan matching method based on situation characteristics

And during a received task, automatically extracting situation characteristics from a superior task instruction and an emergency situation, establishing a current task situation characteristic vector, automatically performing characteristic matching with each plan in a plan library, and calculating the situation similarity of the plans.

(1) Current context feature data acquisition

Connotation: the method senses the current task requirement characteristics and task situation characteristics and is a trigger link for recommending the whole plan.

The acquisition method comprises the following steps:

(1) task characteristics: the user issues tasks in a voice command or a message command mode. The system adopts the methods of voice recognition, natural language processing and knowledge extraction to extract the task requirement information from the natural language description. Taking a target fire attack type task in a military combat application scene as an example, the language description of 'recommending an XX target attack plan and requiring a sink within 2 hours' can identify that the combat style is 'fire attack', the attack target is 'XX number', the damage requirement is 'sink', the completion time limit is '2 hours', and the target type is 'missile destroyer' from a related database.

(2) Situation characteristics: and reading from the situation data. For example, target location data (e.g., latitude and longitude coordinates) readable by the current target location of the target; target state readable target state data (e.g., good/light/heavy); day and night distinguishing can inquire a morning and evening schedule according to longitude and latitude, and compare and judge with the current time; meteorological conditions may read meteorological data (e.g., fly-by/complex fly-by/no-fly).

(2) Plan context similarity calculation

With reference to fig. 2, the method for matching similar plans based on contextual characteristics mainly includes the following steps:

for each protocol: calling a corresponding situation characteristic similarity calculation method for each situation characteristic of the plan, and calculating the similarity of the characteristic; and according to the situation characteristic weight system model, carrying out weighted summation on the similarity of all the characteristics to obtain the task situation similarity of the pre-plan.

And forming a plan similarity data set by the internal codes and the similarity values of all the plans.

Similarity algorithm of each feature:

1) Hit target similarity: string matching is used by the name of the object, e.g., "XX".

2) Target type similarity: refer to the resource classification system. Taking target fire striking tasks in military combat application scenes as an example, the major categories can be divided into sea-based platforms and empty-based platforms, the minor categories are submarines and the like, and the models are different from one another. Similarity of the same type target is 100%, similarity of the same fine type target is 66%, similarity of the same large type target is 33%, and similarity of the different large type targets is 0%

3) The damage requirement is as follows: the current characteristic is X, the plan characteristic is Y, and the value is referred to table 1. Then the damage requires similarity = exp (— X-Y/2).

Table 1 damage requirement value-taking table

Impact and sink	Heavy wound	Minor injury
			1	2	3

4) Completion time limit similarity:

(1) if the scheduled task completion time limit < the current task completion time limit, similarity =1

(2) else similarity = exp (- | scheduled task completion deadline-current task completion deadline |/current task completion deadline).

5) Target position similarity:

(1) calculating the average distance D from the striking target position in the plan to each force starting point position in the plan ₀ ；

(2) Calculating the average distance D from the current striking target position to each force starting point position in the plan ₁ ；

(3) Target position similarity = (D) ₀ -| D ₀ - D ₁ |)/ D ₀ And if the result is less than 0, it is counted as 0.

6) Target prediction position similarity: as above.

7) Target state: the current characteristic is X, the plan characteristic is Y, and the value is referred to the table 2. The target state similarity = exp (— X-Y |/2).

TABLE 2 target State evaluation Table

Intact	Minor injury	Heavy wound
			1	2	3

8) Day and night distinguishing: and (4) calculating a Boolean value.

Example 3: comprehensive multi-factor plan screening and sorting method

And (3) calculating a comprehensive recommendation index by integrating factors in the aspects of situation similarity, resource availability and effect superiority and inferiority, automatically screening and sequencing the plans in the plan library, and generating a plan recommendation list on a human-computer interaction interface.

(1) Force resource availability calculation

With reference to fig. 3, the military resource availability computing method mainly comprises the following steps:

and selecting the resources within a certain radius range around the target to form a peripheral resource subset. The screening resources are calculated according to the position of each resource, the radius of each resource is calculated according to the value of the 'finishing time limit' characteristic in the current task situation characteristic and the motion performance parameter of each resource.

For each plan i, extracting all resources in the content characteristics of the plan, and comparing the resources with each resource in the peripheral available resource subset, wherein the method comprises the following steps of: the types of the calculation force resources are similar. Reference is made to a resource classification system. For example, the similarity of the same type target is 100%, the similarity of the same fine type target is 66%, the similarity of the same large type target is 33%, and the similarity of the different large type target is 0%. The highest similarity among all available resources is selected as resource availability.

Availability of the plan = sum of availability of all resources/total number of resources.

And forming a plan availability data set by the inner codes and the availability of all plans.

(2) Calculation of superiority and inferiority of effect evaluation

With reference to fig. 4, the method for calculating the superiority and inferiority of the effect evaluation mainly includes the following steps:

for each similar and available plan, performing single evaluation on the execution effect, the time consumption and the resource consumption respectively, taking the execution effect as an example, performing normalization processing according to the following method, and the other characteristic algorithms are the same: finding out the maximum value and the minimum value of the execution effect in all similar and available plans; and calculating the execution effect of the plan according to a formula.

(2) And weighting and summing the execution results according to the weight system model to obtain the calculation result of the preplan superiority and inferiority.

(3) And forming a project superiority and inferiority data set by the inner codes and the superiority and inferiority of all the projects.

(3) Plan sort calculation

With reference to fig. 5, the calculation method for the predetermined plan ranking mainly includes the following steps:

1) And for each recommended plan, carrying out weighted summation on the calculation results of the plans in the aspects of similarity, availability and superiority and inferiority 3 according to the comprehensive recommended weight system model to obtain the recommended index of each plan.

2) And (4) sequencing the plans in the subset of the recommended plan library from high to low according to the recommended indexes to obtain the serial number of each plan, and forming a comprehensive sequencing data set of the plans by the serial number and the calculation results of the internal codes, the similarity, the availability and the superiority and inferiority of the plans.

Example 4: method for constructing focus-focused model and optimizing iterative learning

After browsing most recommended plans, the user selects one plan which best meets the requirement and approves the plan by clicking, and the rest plans in the recommendation list do not carry out any operation. Therefore, according to the selection operation of the user, all recommended plans are divided into selected plans and unselected plans. A machine learning method can be used for analyzing a large number of selection behaviors under the same type of task pattern, and the focus of attention of the type of task pattern is obtained.

As described above, for each dimension of the situation feature, the resource usage feature, and the effect evaluation feature, there is a weight value, and the weight data of 3 indexes, namely, the situation similarity, the resource availability, and the effect superiority and inferiority, constitutes a focus model. Initialization may set all weights to an evenly divided mode and then iteratively learn the optimization model by collecting user selection operations.

With reference to fig. 6, the learning optimization method of the focus-focused model mainly includes three steps: calculating the feature difference, correcting the feature difference and correcting the weight of the preset features. We define the evaluation value of each dimension feature of the selected plan as X, and the evaluation value of each dimension feature of the non-selected plan as X

And N is the number of the non-selected cases. In addition, the weight of all features is denoted by W. The purpose of the intelligent learning algorithm is to output the corrected feature weights based on the inputs X, Y and W.

(1) Feature difference calculation

The feature difference is the difference between each feature of the selected plan and the unselected plan, and implies the preference and style of the user. Because there are a plurality of non-selected plans and thus there are a plurality of groups of similarity, we average these similarity according to dimensionality to obtain an overall non-selected plan similarity vector

Namely: />

(2) Feature variance correction

In fact, the number of unselected plans is generally large, and the values of some features may be wide and distributed without concentration, which results in an unobvious tendency of the features, thereby causing a deviation in the adjustment of the weight. And the variance of the group of data with the larger divergence of the distribution is also larger. Therefore, for more precise weight adjustment, the variance of the similarity value is integrated into the feature difference, and the influence of the arrays is weakened through the variance. Specifically, we first calculate the variance σ of the non-selected plan in each feature dimension:

(3) Plan feature weight correction

Through difference correction, the difference based on positive and negative samples is obtained, and a feedback mechanism can be used for adjusting the feature weight:

wherein, α and β are scaling and are mainly used for representing the learning speed of the recommendation system. When these two values are large, the learning speed is fast, but the learning is rough, resulting in a poor final model. When the two values are smaller, learning becomes finer, and the final model effect is better. However, the learning speed is slow, which increases the demand for the number of data sets, i.e., the number of uses by the user. Through the above, the correction of the characteristic weight of the plan by the feedback mechanism is realized by using the difference of the plan. Finally, weight normalization operation is carried out, and then one-time learning is completed.

Claims

1. A method for learning and focusing on optimal sequencing of context-aware plans is characterized by comprising the following steps:

(2) Initializing and constructing a case selection focus model W;

(3) Extracting situation characteristics from a superior task instruction T and an emergency situation S when a subject emergency treatment task is performed, and constructing a current situation characteristic vector SC;

2. The method according to claim 1, wherein the step (1) specifically comprises:

(1.2) for each protocol P in the library P of protocols _i According to the task style M to which it belongs _k According to the corresponding plan feature vector model MM _k Respectively MMS therein _k 、MMC _k 、MME _k Carrying out assignment and constructing the plan P _i Is predetermined characteristic vector PM _i The method specifically comprises the following steps:

(1.2.2) for each MMC _kj According to the predetermined plan P respectively _i Assigning data of each specific type of resource usage to form PMC _i ；

3. The method according to claim 1, wherein the step (2) is specifically as follows:

；

(2.2) all task styles M _k Corresponding case selection focus feature model W _k Forming a set W.

4. The method according to claim 1, wherein the step (3) is specifically as follows:

5. The method for learning context-aware plan preference ranking of focus of interest according to claim 1, wherein the step (4) is specifically:

(4.2.2) calculation of the plan PM _ki Is measured by a condition similarity value Sim _i For all characteristic variables MMS _kj The situation similarity value SS _ij Weighted summation:

；

；

(4.4.1.1) finding PM _k Middle MME _kj Maximum value of

And minimum value

And PM _ki The value of (D) V;

(4.4.1.2) calculating MME _kj Is taken as the value EF _ij ：

；

；

6. The method according to claim 1, wherein the step (5) specifically comprises:

；

7. The method according to claim 1, wherein the step (6) is specifically as follows:

8. The method according to claim 1, wherein the step (7) is specifically as follows:

(7.3) for reference protocol B, SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i Combining to form a reference plan evaluation vector X;

(7.4) forming a set RaM of the N recommended plans ranked in the front in the plan ranking list Ra, which are not selected as reference plans, for each of which the RaM _i ，SS _ij 、FA _ij 、EF _ij 、Sim _i 、Ava _i 、Eff _i Combined to form a predetermined RaM _i Is estimated as vector Y _i Combining the evaluation vectors of all RaMs to form a recommended plan evaluation matrix

。

9. The method according to claim 1, wherein the step (8) is specifically as follows:

：

；

(8.2) calculating the variance of Y in each feature dimension

：

；

(8.3) definition

And

two scaling factors are used for adjusting the learning speed, and the iterative tuning optimization scheme focuses on the key model W:

。

10. a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method of learning context-aware protocol preference ordering for focus of interest as claimed in any one of claims 1-9.

11. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the computer program implements a method of learning context-aware protocol preference ordering for focus of interest as claimed in any one of claims 1 to 9.