CN113159330A - Professional learning path system and method based on hierarchical task network planning model learning - Google Patents

Abstract

The invention discloses a professional learning path system based on hierarchical task network planning model learning, which comprises a Hierarchical Task Network (HTN) model with a professional knowledge structure, a knowledge structure HTN model learning module (HTNML) and a learning path generation module (SHOP 2). The invention relates to the technical field of artificial intelligence planning and online education, and particularly provides a professional learning path system and a method based on the learning of a hierarchical task network planning model.

Description

Professional learning path system and method based on hierarchical task network planning model learning

Technical Field

The invention relates to the technical field of artificial intelligence planning and online education, in particular to a professional learning path system and a method based on hierarchical task network planning model learning.

Background

Personalized learning path recommendation is an important function of online education systems. The learning path recommendation method based on the hierarchical task network planning model learning is characterized in that a learning path is abstractly understood as a hierarchical task network planning model, a professional learning hierarchical task network planning model is constructed from observation of incomplete online education state data, and a proper learning path is calculated according to a learning target of an online education system user and recommended to the user.

In recent years, with the prevalence of online education in various industries at all levels of the whole society, learning path recommendation has been widely researched, and various learning path recommendation methods and systems are endlessly developed.

The method is based on modeling and evaluating the personal knowledge graph of a user, and recommending a learning path to the user according to an evaluation result, the method relies on modeling the knowledge graph of the personal knowledge structure of the user, the accuracy and the effectiveness of the method heavily depend on the degree of mastering the integrity of the knowledge structure of the user, and the method is lack of understanding and application on the objective professional knowledge structure and the internal logic; a course learning path recommendation method and device based on user behavior data analysis are provided by Wang, Zhang Fuchun and the like, the method is used for mining associated weights among learning courses from user behavior (online course repairing and reading) data, and recommending learning paths to users according to the weights, the method depends on the existing user data of a system, and recommends the learning paths according to the user data, has certain objectivity, but the objectivity can not comprehensively reflect the objectivity of the internal logic of knowledge, appears narrower, and lacks of analysis on the personalized requirements of the users; the method can embody the objective internal logic of a professional knowledge structure and can also carry out dynamic recommendation according to the learning track of the user, but the maintenance of the knowledge point network is a larger knowledge engineering, and the maintenance difficulty and the cost are larger; zhang Jia Lei provides a personalized learning path planning method, the method combines knowledge map and user's learning state to carry on the dynamic learning path recommendation, in addition, the method relies on the knowledge point structure in the teaching material, more suitable for the standardized education of middle and primary schools, it is difficult to reflect the level differentiation and the inherent logic of the professional knowledge point structure to the professional education, therefore is not suitable for professional learning of professional; the application-driven learning path recommendation method based on the data map, the information map and the knowledge map, which is proposed by segmentally, Zhan-zhao, Cao Kai and the like, depends on the construction of various knowledge maps, and recommends learning points and learning paths by obtaining the learning dynamics of a user, so that the method also faces the problems that the knowledge map is difficult to automatically evolve, so that the maintenance is difficult and the like; the method is characterized in that a mathematical model of a learning path optimization problem is established, and a lightning search algorithm is used for solving the problem to obtain an optimal learning path.

Hierarchical Task Network (HTN) planning is an artificial intelligence planning (AI planning) method. In a given initial state, the method considers the process of achieving certain targets as the completion process of one task, and the tasks can be represented as a hierarchical structure, namely: a task may continue to be subdivided into a series of sub-tasks that are completed in sequence, one of the most primitive sub-tasks being called an action. Atomic activities where actions are not separable, actions are defined by action names, parameters, preconditions and effects for action execution. The preconditions and effects of an action can in turn be defined as a set of logical propositions. Under the framework of HTN planning, the learning of each course is defined as an action, a series of actions can be combined into a subtask, and the purpose of professional knowledge is expressed as a task. This gives a good description of the hierarchy of expertise learning, as well as the inherent logic between courses.

One of the difficulties in programming with HTN is that HTN modeling is a complicated knowledge engineering and is difficult to update and maintain. In order to solve the problem, the invention provides a method for obtaining a complete professional knowledge HTN model by using an HTNML learning tool and observing data generated in user learning, and a learning path is calculated by using an HTN planning tool such as SHOP2 and the like according to a learning target input by a user.

Disclosure of Invention

Aiming at the situation and overcoming the technical defects at present, the invention provides a professional learning path system and a method based on the hierarchical task network planning model learning, the method abstracts and understands the learning path as the hierarchical task network planning model, starts from observing incomplete online education state data (namely user data) on the basis of an incomplete professional knowledge structure HTN model constructed by experts, constructs a professional learning hierarchical task network planning model, and calculates a proper learning path to recommend to a user according to the learning target and the requirement of the online education system user.

The technical scheme adopted by the invention is as follows: the invention relates to a professional learning path system based on hierarchical task network planning model learning, which comprises a Hierarchical Task Network (HTN) model of a professional knowledge structure, a knowledge structure HTN model learning module (HTNML) and a learning path generation module (SHOP2), wherein the Hierarchical Task Network (HTN) model planning problem is defined as a quadruple: (S)₀，T，M，A)，S₀Is an initial state, which is a set of propositions; t is a list of tasks that need to be completed; m is a set of methods that illustrate how a high level task is decomposed into a fully ordered set of low level sub-tasks; a is a set of action sets corresponding to the original subtasks that can be directly executed; in the above definition, each task has a unique task name and has zero or more parameters, each parameter is a variable symbol or a constant symbol, and the method is defined as (m, t, Pre, Sub), where m is a unique method name and has zero or more parameters; t is the head task of the method decomposition; pre is a precondition for the implementation of the method; sub is a subtask list that shows how the head task is decomposed into a series of subtasks; the m parameter is composed of a parameter of the head task, a parameter of each subtask, and all parameters used in the preconditionWhich may be an atomic task (i.e., an action), or a non-atomic task (i.e., must be further decomposed), the prerequisites for each method are represented as a set of logical words that represent a set of prerequisites that must be satisfied before the method can be applied; the knowledge structure HTN model learning module takes an existing incomplete professional knowledge structure HTN model and incompletely observed online education state data (namely user data) as input, and applies an existing HTNML system to construct a professional learning hierarchical task network planning model, so that the problem of how to obtain a more complete HTN model under the conditions that the HTN model of the knowledge structure is incomplete and the user data is incompletely observable is solved; the learning path generation module is used for calculating a learning path and recommending the learning path to the user on the premise that the user gives a learning target and a learning requirement based on the learned knowledge structure HTN model.

Further, the algorithm flow of the knowledge structure HTN model learning module is specifically as follows: based on an incomplete HTN model, an existing HTNML system is applied to input a data set, a predicate list, an action model list and a method list are extracted from observed user data, state constraints, action constraints and decomposition constraints are respectively constructed, the constraints are given to weights and input into an MAX-SAT tool to be solved, a more complete action model and a more complete precondition of the method are obtained, and finally the more complete HTN model is formed.

Further, the action model comprises action parameters, preconditions and effect descriptions.

Further, the learning path generation module uses an existing HTN planner such as SHOP2 to construct an HTN planning problem, i.e., a problemm file, based on the learning objective of the user, and the SHOP2 can solve the learning path according to the input model file and the problem file.

The method for the professional learning path based on the hierarchical task network planning model learning specifically comprises the following steps:

the method comprises the following steps: the system configuration, before using HTNML (hierarchical task network model learning system), the configuration file is needed, the directory of the first behavior data set, the number of the second behavior data set, the weight occupied by observation, the proportion occupied by incomplete description of the decomposition tree in the data set are configured in the configuration file, the MAX-SAT solver file is respectively configured in the fifth line, and the address of the output result is configured in the last line;

step two: extracting an HTN model, wherein the HTN model is extracted in the step, and comprises a prediction list, an action mode list and a method structure list; firstly, scanning all decomposition trees, and replacing all objects with corresponding variables, wherein each variable is constrained by one type; secondly, collecting all predicate composition predicate lists, for example, (on x y) becomes (onx; then all different action model families are used as an action model list; finally, all different tasks are decomposed, and each task is composed of corresponding subtasks and serves as a method structure list;

step three: establishing a state constraint, wherein in the decomposition tree, if a predicate frequently appears before the predicate executes the action and the parameter of the predicate is also the parameter of the action, the predicate is very likely to be a precondition of the action, and similarly, if the predicate frequently appears before the method is applied, the predicate is very likely to be a prerequisite of the method; if a predicate occurs frequently after the action constraint operation is performed, it is likely to be an action effect; in the process we learn, this information will be encoded in the form of constraints, which we call state constraints since they are established by the relationships between states and actions, states and methods;

step four: establishing a decomposition constraint, in which the decomposition constraint encodes the structural information provided by the decomposition tree, if task T can be decomposed into N subtasks ST1, ST2, …, STn, the subtask STi providing some prerequisites for the method of subtask STi +1, so that the method can apply the next subtask STi +1, and furthermore, the parameters of the prerequisites (or effects) should be included in the parameters of the action or method to which the preconditions (effects) belong;

step five: the action constraints are established to ensure that the learned action model is effective and repeatable, can reflect the action model of certain characteristics of the real world, and further induce some constraints in different action categories, wherein the constraints act on the inseparable personal behaviors and can be divided into the following two types: (1) before applying the action, the action may not add the fact that it is already true; (2) an action cannot delete the fact that it did not exist before the application action; these constraints were placed previously to ensure that the learned action models were compact and that most existing planning domains satisfy them. However, our learning algorithms are perfect without them;

step six: solving the constraints, a prerequisite for encoding motion model information and methods by establishing three constraints, the relative importance of which must be determined before the constraints are solved by weighting Max-set, for which three new parameters B are introduced_i(1<i<3) Controlling the weight of the constraint by controlling the parameters according to a formula

By simple adjustment of B_iThe weight value can be adjusted to be 0 to infinity, and finally, a constraint solution is calculated through MAX-SAT to form the HTN model.

The invention with the structure has the following beneficial effects: the invention relates to a professional learning path system and a method based on hierarchical task network planning model learning, which can better reflect the objective internal logic of professional knowledge, simultaneously give consideration to the application of user data and the professional knowledge requirement of a user, and can ensure that the system maintenance is better, and the invention has the other advantages that a knowledge structure HTN model learning module (HTNML) is used, the module uses an HTN learning tool, namely HTNML, and can obtain a more perfect HTN model by observing the learning effect of each course of the user under the condition of giving an incomplete HTN model, and a proper learning path can be calculated by using an HTN planning tool, namely SHOP2 by inputting the professional learning target of the user under the more perfect HTN model, and the learning path is abstractly understood as a hierarchical task network planning model, and starting from observation of incomplete online education state data (namely user data), constructing a professional learning level task network planning model, and calculating a proper learning path according to the learning target and the requirement of the online education system user and recommending the learning path to the user.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:

FIG. 1 is a block diagram of a system and method for a professional learning path based on hierarchical task network planning model learning according to the present invention;

FIG. 2 is a state constraint diagram of the professional learning path system and method of the present invention based on hierarchical task network planning model learning;

FIG. 3 is a Hierarchical Task Network (HTN) model structure diagram of a professional knowledge structure of a hierarchical task network planning model learning-based professional learning path system and method of the present invention, which is exemplified by an algorithm direction software engineer;

FIG. 4 is a block diagram of an algorithm flow of a professional learning path system and method knowledge structure HTN model learning module based on hierarchical task network planning model learning in accordance with the present invention;

fig. 5 is a schematic diagram of a learning path generation module (SHOP2) taking an action model for training as an example in the professional learning path system and method based on hierarchical task network planning model learning according to the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments; all other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that the terms "front," "back," "left," "right," "upper" and "lower" used in the following description refer to directions in the drawings, and the terms "inner" and "outer" refer to directions toward and away from, respectively, the geometric center of a particular component.

FIG. 3 schematically illustrates the structure of HTN knowledge required by an algorithm-oriented software engineer and courses required to be reviewed, and taking the knowledge structure of the algorithm-oriented software engineer as an example shown in FIG. 3, the HTN model of the expert knowledge structure is defined as (S)₀T, M, a), wherein: s₀＝{(finished highSchool)(clear cpgm),(clear amath),(clear lagb),(clear pst),(clear ds),(clear adn)}

Wherein: (refined high school) means that high school stage learning is completed, clear predicate expression does not have some knowledge, cpgm means C language programming knowledge, amath means high mathematics knowledge, lagb means linear algebra knowledge, pst means probability theory and mathematical statistics knowledge, ds means data structure knowledge, and adn means algorithm design;

T＝{AcquireProgrammingSkills,AcquireMathBasics,AcquireAlgorithmBasics,AcquireEngineeringBasics}

respectively representing the need of having a programming basis, a mathematical basis, an algorithm basis and an engineering basis;

m ═ method programing skills, method mathbases, method engineering basics } exemplifies method programing skills, which are defined as (method programing skills, language c, constr, Sub), Pre { (find high school) }, Sub ═ language c, language cpp }, where language c and language cpp are atomic tasks (actions), respectively, defined as < name, precondition, ADD, DEL >. Taking languageC as an example, name { (refined high school) }, ADD { (acquire program bases), (acquire Cprogramming) }, and DEL { }.

Taking the algorithm engineer in fig. 3 as an example, the proplem file contains the following contents:

(defproblem problem ALG

((clear c)(clear am)(clear la)(clear ps)(clear ds)(clear ad)(clear

uml)(clear se)(clear st)(clear bofca)(clear daac)(clear ed)(clear

nb)(clear ns))

((Alogrithm_engineer c am la ps ds ad))

FIG. 4 is a block diagram of an algorithm flow of a professional learning path system and method knowledge structure HTN model learning module based on hierarchical task network planning model learning, a predicate list, an action model list and a method list are extracted from an observation data set, a state constraint, an action constraint and a decomposition constraint are respectively constructed and combined, the constraint sets are input to an MAX-SAT solver together, and preconditions and effect predicates of actions and preconditions of the method are obtained through calculation, so that a complete HTN model is formed finally.

Fig. 5 is a schematic diagram of the learning path generation module (SHOP2) taking the motion model for training as an example, training the motion model by inputting the number of data sets respectively 10, 100, and 200, and comparing the fitting degree of the course recommendation model generated by HTNML motion model learning with the course recommendation model designed by me. FIG. 5 is an example of planning an algorithmic engineer course learning route that has been trained to have a course recommendation model restored by the JSHOP2 planner for a 200 dataset size.

In summary, the action model learning method based on machine learning of the present invention obtains the user personalized learning model from the user learning data, and proposes a feasible personalized learning path for the user based on the user personalized learning model for the user target.

It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (5)

1. The professional learning path system based on the hierarchical task network planning model learning is characterized by comprising a hierarchical task network model of a professional knowledge structure, a knowledge structure HTN model learning module and a learning path generation module, wherein the hierarchical task network model planning problem is defined as a quadruple: s₀，T，M，A，S₀Is an initial state, which is a set of propositions; t is a list of tasks that need to be completed; m is a set of methods that illustrate how a high level task is decomposed into a fully ordered set of low level sub-tasks; a is a set of action sets corresponding to the original subtasks that can be directly executed; in the above definition, each task has a unique task name and has zero or more parameters, each parameter is a variable symbol or a constant symbol, and a method is defined as m, t, Pre, Sub, m is a unique method name and has zero or more parameters; t is the head task of the method decomposition; pre is a precondition for the implementation of the method; sub is a subtask list that shows how the head task is decomposed into a series of subtasks; the m parameters are composed of parameters of a head task, parameters of each subtask and all parameters used in the precondition, each subtask can be an atomic task or a non-atomic task, the precondition of each method is represented as a group of logic words and represents a set of preconditions which must be met before the method is applied; the knowledge structure HTN model learning module takes an existing incomplete professional knowledge structure HTN model and incompletely observed online education state data as input, and a professional learning hierarchical task network planning model is constructed by applying an existing HTNML system; the learning path generation module is used for calculating a learning path and recommending the learning path to the user on the premise that the user gives a learning target and a learning requirement based on the learned knowledge structure HTN model.

2. The professional learning path system based on hierarchical task network planning model learning of claim 1, wherein the algorithm flow of the knowledge structure HTN model learning module is specifically as follows: based on an incomplete HTN model, an existing HTNML system is applied to input a data set, a predicate list, an action model list and a method list are extracted from observed user data, state constraints, action constraints and decomposition constraints are respectively constructed, the constraints are given to weights and input into an MAX-SAT tool to be solved, a more complete action model and a more complete precondition of the method are obtained, and finally the more complete HTN model is formed.

3. The system of claim 2, wherein the action model comprises action parameters, preconditions, and effect descriptions.

4. The system of claim 1, wherein the learning path generation module uses an existing HTN planner such as SHOP2 to construct an HTN planning problem, i.e. a problemm file, based on the learning objective of the user, and SHOP2 can solve the learning path according to the input model file and problem file.

5. The method for the professional learning path based on the hierarchical task network planning model learning is characterized by comprising the following steps:

the method comprises the following steps: the method comprises the steps of system configuration, before a hierarchical task network model learning system is used, Profile files need to be configured, a directory of a first behavior data set, the number of a second behavior data set, a third behavior data set are weights occupied by observation, a fourth behavior data set is a proportion occupied by incomplete description of a decomposition tree in the data set, a fifth behavior data set is an MAX-SAT solver file, and a last behavior data set is an address of an output result;

step two: extracting an HTN model, wherein the HTN model is extracted in the step, and comprises a prediction list, an action mode list and a method structure list; firstly, scanning all decomposition trees, and replacing all objects with corresponding variables, wherein each variable is constrained by one type; secondly, collecting all predicates to form a predicate list and changing the predicate list into a predicate list; then all different action model families are used as an action model list; finally, all different tasks are decomposed, and each task is composed of corresponding subtasks and serves as a method structure list;

step three: establishing a state constraint, wherein in the decomposition tree, if a predicate frequently appears before the predicate executes the action and the parameter of the predicate is also the parameter of the action, the predicate is very likely to be a precondition of the action, and similarly, if the predicate frequently appears before the method is applied, the predicate is very likely to be a prerequisite of the method; if a predicate occurs frequently after the action constraint operation is performed, it is likely to be an action effect;

step four: establishing a decomposition constraint, in which the decomposition constraint encodes the structural information provided by the decomposition tree, if task T can be decomposed into N subtasks ST1, ST2, …, STn, subtask STi providing some prerequisites for the method of subtask STi +1, the parameters of the prerequisites or effects should be included in the parameters of the action or method to which the preconditions or effects belong;

step five: the action constraints are established to ensure that the learned action model is effective and repeatable, can reflect the action model of certain characteristics of the real world, and further induce some constraints in different action categories, wherein the constraints act on the inseparable personal behaviors and can be divided into the following two types: (1) before applying the action, the action may not add the fact that it is already true; (2) an action cannot delete the fact that it did not exist before the application action;

step six: solving the constraints, a prerequisite for encoding motion model information and methods by establishing three constraints, the relative importance of which must be determined before the constraints are solved by weighting Max-set, for which three new parameters B are introduced_i(1<i<3) Controlling the weight of the constraint by controlling the parameter according toFormula (II)

By simple adjustment of B_iThe weight value can be adjusted to be 0 to infinity, and finally, a constraint solution is calculated through MAX-SAT to form the HTN model.

Images