Functional Knowledge Integration Method for Innovative Design Based on Deep Reinforcement Learning

Abstract

The functional knowledge integration method for innovative design based on deep reinforcement learning includes: 1) constructing a functional knowledge graph by matching discrete functional knowledge units with their input / output; 2) modeling the core task of functional knowledge integration—functional unit chain generation—as a Markov decision process and solving it using the deep Monte Carlo algorithm in reinforcement learning; 3) merging the phased solutions (functional unit chains) generated by the reinforcement learning algorithm into a final design solution through an automatic or semi-automatic iterative framework. Based on the user's structured semantic description of the target product's functions, this invention can output the most functionally complete design solution in near real-time, ensuring solution quality while significantly reducing the time cost required for functional knowledge integration under a large-scale knowledge base.

Description

Technical Field

[0001] This invention relates to the fields of deep reinforcement learning, conceptual design, and functional knowledge integration. First, the functions of the product to be designed are converted into structured semantic information—target input / output. Then, based on a deep Monte Carlo algorithm, a mapping from the target input / output to the optimal conceptual design scheme is learned. This transforms the functional knowledge integration process from a search to a Markov decision process, significantly reducing the time cost required for functional knowledge integration, especially in large-scale functional knowledge bases. Background Technology

[0002] The complexity of design tasks is increasing daily, and the knowledge required for many product designs is shifting from single-disciplinary to multi-disciplinary. This trend has not only led to more general knowledge representation models but also significantly increased the size of design knowledge bases. Traditional functional unit integration (FKI) algorithms, which rely on traversal, suffer from combinatorial explosion in large search spaces, making them largely impractical. Therefore, it is essential to explore a more efficient algorithm with a completely different core to effectively guarantee the feasibility of FKI under large-scale knowledge bases. Summary of the Invention

[0003] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide a functional knowledge integration method for innovative design based on deep reinforcement learning.

[0004] This invention generates functional unit chains in real time by replacing the traditional traversal search algorithm with a deep reinforcement learning algorithm, thus solving the problems of low feasibility and high running time cost of previous algorithms under large-scale knowledge bases.

[0005] This invention relates to a functional knowledge integration method for innovative design based on deep reinforcement learning, comprising the following steps:

[0006] 1) Construct a functional knowledge graph by relying on the matching relationship between discrete functional knowledge units and their inputs / outputs.

[0007] 2) The core task of functional knowledge integration—the generation of functional unit chains—is modeled as a Markov decision process and solved using the deep Monte Carlo algorithm in reinforcement learning.

[0008] 3) The phased solutions (functional unit chains) generated by the reinforcement learning algorithm are merged into the final design solution through an automatic or semi-automatic iterative framework. This achieves an efficient FKI framework that takes the user's structured semantic description of the target function as input and the corresponding functional design solution as output.

[0009] Furthermore, step 1) specifically includes:

[0010] 1.1) Construct a functional knowledge graph (FKG) using the smallest reusable unit of functional design knowledge—the functional unit—and represent it as a directed graph G(V,E). Each node v... i ∈V represents a functional unit, and each edge e ij ∈E represents v i and v j The matching relationships between them. Knowledge in a knowledge graph uses triples (v... i e ij v j The format exists, specifically representing the functional unit v. i The output is controlled by the functional unit v j The inputs that match each other are called predecessors and successors;

[0011] 1.2) The ontology of a functional unit is defined as having the following seven attributes: ID, Input, Output, Prior, Next, Category, and Carrier. ID is the unique identifier of the functional unit. Input and Output represent the input and output of the functional unit, describing the basic content of the product function. Prior and Next reflect the position (or adjacency) of the functional unit in the FKG, corresponding to predecessor and successor, respectively. Category refers to the functional category carried by the functional unit, including four types: transformation, support, storage, and activation functions. Carrier represents the physical or technical prototype of the functional unit in reality;

[0012] 1.3) The representation of inputs and outputs is the key to functional representation. This invention adopts a keyword plus parameter representation method. Specifically, each keyword exists in the format of "modifier + core word", such as "fresh, clean + water". Each parameter exists in the format of "name + value range + unit", such as "voltage +

[220] + V". Each input or output can be composed of multiple keywords and parameters. Each keyword can have multiple modifiers, but can only correspond to one core word.

[0013] 1.4) To save space and resolve ambiguity, this invention designs a bucket mechanism for polysemous word management. Specifically, each word is assigned to a different bucket based on its meaning, and the bucket is uniformly represented by a tag. Therefore, for any polysemous word appearing during the functional unit input process, the user can determine its specific meaning by assigning it to different buckets.

[0014] 1.5) The matching relationship between functional units is defined as follows: For functional unit v i and v j If vi The arbitrary output of v j If any input matches, then it is called a functional unit v. i and v j Yes, they are matched. For a pair of inputs and outputs, they are considered a match only if their core terms are in the same bucket, and the modifiers and features of the output include those of the input (specifically, if a parameter appears in both input and output, its value range in the input should be included in the value range of the output). The number of unmatched inputs / outputs between adjacent functional units, as a quantitative indicator of the matching relationship, exists as a label on each edge of the functional knowledge graph and is called the redundancy number.

[0015] Furthermore, step 2) specifically includes:

[0016] 2.1) Represent the target function as two virtual nodes St and Ed in the Functional Keying Group (FKG), where the former contains only output attributes and the latter contains only input attributes. Then, the FKI task can be defined as finding a completely matching path from St to Ed in the FKG, i.e., the sum of redundancies between functional units in the path is 0. Further utilizing the idea of ​​dynamic programming, the complex FKI task can be transformed into an iterative process of searching a partially matching path (i.e., a chain of functional units) from St to Ed each time, and then treating the unmatched inputs / outputs as new target functions and searching their corresponding functional unit chains, gradually completing the matching of all inputs / outputs.

[0017] 2.2) The sum of redundancies in the path is used as an evaluation metric for the functional unit chain, called global redundancy. By optimizing this metric, both the functional completeness of the obtained scheme and the number of iterations required by the algorithm can be controlled within a reasonable range.

[0018] 2.3) The FKI task is modeled as a Markov decision process, where the behavior at each time step corresponds to the selection of a successor node in the path search process.

[0019] 2.4) Construct a value network Q-network to make decisions about all successor nodes starting from the target input node St. The network includes two feature extraction modules: convolutional and LSTM, and an MLP prediction module. It receives seven pieces of information as input: the current node, neighboring nodes, the redundancy between the current node and its successor nodes, the behavioral trajectory, the global redundancy, and the target functional requirements. The output is the expected reward value Q-value for each successor node.

[0020] 2.5) Rewards are used to guide the agent's behavior, and the definition of the reward function is a crucial part of reinforcement learning algorithm design. This invention defines four reward functions: success, failure, progress, and setback. These provide feedback for four scenarios: successfully reaching the target output node, reaching the maximum search length but not yet reaching the final node, being at an intermediate node, and being at an intermediate state with no successor node. The failure reward is a non-zero reward based on the health-oriented hypothesis, enabling the agent to learn how to perform a healthy search even when the search fails (even if the path redundancy decreases). This reward significantly improves the algorithm's convergence speed and performance.

[0021] 2.6) The value network is trained using the Every-visit Monte Carlo algorithm. The target policy is defined as a greedy policy, and the behavior policy is an improved epsilon greedy policy with an interval epsilon greedy policy whose epsilon value decreases during the training process.

[0022] Furthermore, step 3) specifically includes:

[0023] 3.1) The structured description of the target function input by the user is used as the input to the reinforcement learning algorithm in step 2), and a corresponding optimal functional unit chain is output. Since the FKI process has now been transformed from a search into a Markov decision process, the functional unit chain in this step can be generated at almost real-time speed.

[0024] 3.2) The unmatched inputs / outputs in the functional units generated in step 3.1) are used as the outputs / inputs of the target function in the next iteration. In this step, users are allowed to adjust the existing unmatched inputs / outputs, i.e., switch to a semi-automatic iteration framework. Allowed adjustments include, but are not limited to, deletion, addition, manual matching, and parameter modification.

[0025] 3.3) Iterate through steps 3.1) and 3.2) until all inputs / outputs in the path have been fully matched or confirmed by the user as fully matched (semi-automatic).

[0026] This invention mainly comprises two parts: a functional unit representation / storage method based on buckets and knowledge graphs, and a functional unit chain generation algorithm based on deep reinforcement learning. For the functional unit representation / storage method, firstly, a functional unit ontology is constructed, and its input / output attributes are described using keyword representation. Then, a bucket mechanism is used to uniformly represent synonyms in the keywords. Finally, a knowledge graph is used to store the functional units and their relationships. In the deep reinforcement learning-based functional unit chain generation algorithm, the functional knowledge integration problem is modeled as a Markov decision process. After designing the corresponding value network, strategy, and reward, a deep Monte Carlo algorithm is used to solve the problem, achieving near real-time functional unit chain generation. Based on the user's structured semantic description of the target product's functions, this invention can output the most functionally complete design scheme at almost real-time speed, ensuring the quality of the scheme while significantly reducing the time cost required for functional knowledge integration under a large-scale knowledge base.

[0027] The advantages of this invention are: when describing functions, synonym management of keywords not only simplifies the process of entering functional units but also avoids ambiguity in expression; the process of integrating functional knowledge is transformed from a search process to a decision-making process, and with the help of reinforcement learning algorithms, under a large-scale knowledge base, it can achieve a significant improvement in the speed of solution output and the operability of the algorithm by sacrificing only a small amount of solution output quality. Attached Figure Description

[0028] Figure 1 This is a flowchart of the present invention.

[0029] Figure 2 This is a schematic diagram illustrating an example of the functional unit construction of the present invention.

[0030] Figure 3 This is a schematic diagram of the value network structure of the present invention.

[0031] Figure 4 (a) and Figure 4 (b) is the convergence curve of the reinforcement learning algorithm, which records the changes of the loss value and the average reward value of the last 100 iterations with the number of scenario iterations. The frequency is more than 1 time every 500 scenario iterations.

[0032] Figure 5 This is a schematic diagram illustrating an example of the design scheme generated by the present invention.

[0033] Figure 6 This is a schematic diagram illustrating an example of the generation process of the design scheme of the present invention. Detailed Implementation

[0034] The invention will be further described below with reference to the accompanying drawings.

[0035] This embodiment proposes a mechanical product innovation design method that applies the functional knowledge integration method based on deep reinforcement learning for innovative design according to the present invention, including the following steps:

[0036] 1) Construct a functional knowledge unit ontology, decomposing several mechanical products into mechanical parts, abstracting functional knowledge from them and describing them as functional knowledge units. Then, construct a functional knowledge graph from the discrete functional knowledge units based on the matching relationships between their inputs and outputs.

[0037] 2) The core task of functional knowledge integration—the generation of functional unit chains—is modeled as a Markov decision process. The corresponding value network, strategy, and reward are designed and solved using the deep Monte Carlo algorithm in reinforcement learning.

[0038] 3) When the design requirements of a new mechanical product are submitted to this framework, based on the structured semantic description of its function, a phased solution (functional unit chain) is first generated in real time through reinforcement learning methods. Then, through an automatic or semi-automatic iterative framework, the phased solutions are merged into the final design solution.

[0039] Step 1) proposes a method for representing / storing functional knowledge units based on knowledge graphs, specifically including:

[0040] 1.1) Construct a functional knowledge graph (FKG) using the smallest reusable unit of functional design knowledge—the functional unit—and represent it as a directed graph G(V,E). Each node v... i ∈V represents a functional unit, and each edge e ij ∈E represents v i and v j The matching relationships between them. Knowledge in a knowledge graph uses triples (v... i e ij v j The format exists, specifically representing the functional unit v. i The output is controlled by the functional unit v j The inputs that match each other are called predecessors and successors;

[0041] 1.2) The ontology of a functional unit is defined as having the following seven attributes: ID, Input, Output, Prior, Next, Category, and Carrier. ID is the unique identifier of the functional unit. Input and Output represent the input and output of the functional unit, describing the basic content of the product function. Prior and Next reflect the position (or adjacency) of the functional unit in the FKG, corresponding to predecessor and successor, respectively. Category refers to the functional category carried by the functional unit, including four types: transformation, support, storage, and activation functions. Carrier represents the physical or technical prototype of the functional unit in reality;

[0042] 1.3) The representation of inputs and outputs is the key to functional representation. This invention adopts a keyword plus parameter representation method. Specifically, each keyword exists in the format of "modifier + core word", such as "fresh, clean + water". Each parameter exists in the format of "name + value range + unit", such as "voltage +

[220] + V". Each input or output can be composed of multiple keywords and parameters. Each keyword can have multiple modifiers, but can only correspond to one core word. In this embodiment, the maximum number of modifiers is limited to 5;

[0043] 1.4) To save space and resolve ambiguity, this invention designs a bucket mechanism for polysemous word management. Specifically, each word is assigned to a different bucket based on its meaning, and the bucket is uniformly represented by a tag. Therefore, for any polysemous word appearing during the functional unit input process, the user can assign it to different buckets to determine its specific meaning. The specific construction method of the bucket is as follows:

[0044] (a) When the first word is added to the thesaurus, a bucket will be created for it. This bucket will be assigned a unique identifier tag, the value of which is equal to the first character stored in the bucket by default.

[0045] (b) When a new word is added to the dictionary, the user will decide whether to assign it to one of the existing top n recommended buckets or create a new bucket with a new meaning for it. Recommendations are given based on the cosine similarity between the word and the tag of each bucket.

[0046] 1.5) The matching relationship between functional units is defined as follows: For functional unit v i and v j If v i The arbitrary output of v j If any input matches, then it is called a functional unit v.i and v j Yes, they are matched. For a pair of inputs and outputs, they are considered a match only if their core terms are in the same bucket, and the modifiers and features of the output include those of the input (specifically, if a parameter appears in both input and output, its value range in the input should be included in the value range of the output). The number of unmatched inputs / outputs between adjacent functional units, as a quantitative indicator of the matching relationship, exists as a label on each edge of the functional knowledge graph and is called the redundancy number.

[0047] 1.6) Decompose several mechanical products into mechanical parts, abstract functional knowledge from them, and represent them as functional knowledge units according to the methods described in steps 1.1) to 1.4). (Appendix) Figure 2 The diagram shows an example of functional unit construction. Next, following the method described in step 1.5), the matching relationships between all functional knowledge units are obtained through an automatic calculation program. Finally, each functional unit and its relationships are stored as a functional knowledge graph.

[0048] Step 2) proposes a functional unit chain generation algorithm based on deep reinforcement learning, specifically including:

[0049] 2.1) Represent the target function as two virtual nodes St and Ed in the Functional Keying Group (FKG), where the former contains only output attributes and the latter contains only input attributes. Then, the FKI task can be defined as finding a completely matching path from St to Ed in the FKG, i.e., the sum of redundancies between functional units in the path is 0. Further utilizing the idea of ​​dynamic programming, the complex FKI task can be transformed into an iterative process of searching a partially matching path (i.e., a chain of functional units) from St to Ed each time, and then treating the unmatched inputs / outputs as new target functions and searching their corresponding functional unit chains, gradually completing the matching of all inputs / outputs.

[0050] 2.2) The sum of redundancies in the path is used as an evaluation metric for the functional unit chain, called global redundancy. By optimizing this metric, both the functional completeness of the obtained scheme and the number of iterations required by the algorithm can be controlled within a reasonable range.

[0051] 2.3) The FKI task is modeled as a Markov decision process, where the behavior at each time step corresponds to the selection of a successor node in the path search process;

[0052] 2.4) Constructing a value network (Q-network) for decision-making on all successor nodes starting from the target input node St. The network includes two feature extraction modules: convolutional and LSTM, and an MLP prediction module. It receives seven pieces of information as input: the current node, neighboring nodes, the redundancy between the current node and its successor nodes, the behavioral trajectory, the global redundancy, and the target functional requirements. The output is the expected reward value (Q-value) for each successor node. Specifically, for each input feature, except for the redundancy, which is directly passed as a numerical value, all other text features are represented using corresponding word vectors. To save vector space, this embodiment only encodes the trajectory features of the first five steps and averages the six word vectors for each input / output. The specific structure of the value network is shown in the appendix. Figure 3 The encoding results of the value network input features are shown in the following tabular form of the program block:

[0053]

[0054] 2.5) Rewards are used to guide the agent's behavior, and defining the reward function is a crucial part of reinforcement learning algorithm design. This invention defines four reward functions: success, failure, progress, and obstacle avoidance. These provide feedback for four scenario states: successfully reaching the target output node, reaching the maximum search length but not yet reaching the final node, being at an intermediate node, and being at an intermediate state without a successor node. The failure reward is a non-zero reward based on the health-oriented hypothesis, enabling the agent to learn how to perform healthy searches even when the search fails (even if the path redundancy decreases). This reward significantly improves the algorithm's convergence speed and performance.

[0055] 2.6) The value network is trained using the Every-visit Monte Carlo algorithm. The target policy is defined as a greedy policy, and the behavior policy is an improved epsilon greedy policy with an interval epsilon greedy policy whose epsilon value decreases during the training process. The specific training algorithm is described in the following code block:

[0056]

[0057] In this example, training was completed in 3 days on a small server equipped with a 12th-generation Intel® Core™ i7-12700KF CPU @ 3.61GHz processor and an NVIDIA GeForce RTX 3080 10GB GPU. Figure 4 The image shows the final training convergence curve.

[0058] Step 3) Verify the effectiveness of the method of the present invention based on a real functional knowledge base of mechanical parts, specifically including:

[0059] 3.1) The structured description of the target function input by the user is used as the input to the reinforcement learning algorithm in step 2), and a corresponding optimal functional unit chain is output. Since the FKI process has now been transformed from a search into a Markov decision process, the functional unit chain in this step can be generated at almost real-time speed.

[0060] 3.2) The unmatched inputs / outputs in the functional units generated in step 3.1) are used as the outputs / inputs of the target function in the next iteration. This step allows users to adjust the existing unmatched inputs / outputs, i.e., switch to a semi-automatic iteration framework. Allowed adjustments include, but are not limited to, deletion, addition, manual matching, and parameter modification.

[0061] 3.3) Iterate through steps 3.1) and 3.2) until all inputs / outputs in the path have been fully matched or confirmed by the user as fully matched (semi-automatic), and merge all phased solutions (functional unit chains) into the final design solution;

[0062] 3.4) Based on the design background of establishing a clean and sustainable leisure center, a structured description of the functional design requirements of twenty-one real mechanical products, including generating indoor heating and gas through daily waste treatment, is presented. Based on steps 3.1), 3.2), and 3.3), twenty-one design schemes are generated respectively. For example, in the appendix... Figure 5 In the functional design example shown, a mechanical device is needed to generate electricity to power the entire leisure center. Therefore, its function can be described by the input "flowing river," the output "220V + electricity," and several detailed parameters. Through step 3.1), the first round of iteration results are obtained, i.e., the attached... Figure 6 The functional unit chain is shown in the gray dashed box. The input "speed adjustment signal" of functional unit FU34 is not matched by any output in the chain, and the algorithm automatically determines it as the target output for the next iteration. Furthermore, since there are no unmatched outputs in the chain, "220V + electricity," which is most relevant to the "speed adjustment signal," is manually added as the target input, as described in step 3.2). Following step 3.3), in the second iteration, the algorithm obtains a chain as shown in the attached diagram based on the new target input "220V + electricity," the new target output "speed adjustment signal," and several parameters. Figure 6 The new functional unit chain is shown in the green dashed box. The iteration stops because there are no unmatched inputs or outputs, as shown in the appendix. Figure 6 The blue line indicates the merging of all functional unit chains to the appendix. Figure 5 The final design scheme shown.

[0063] 3.5) Compare the twenty-one design schemes generated in step 3.4) with the pre-set optimal design scheme. The results show that the overlap rate between the design schemes output by the method of the present invention and the optimal design scheme reaches 100%, and the average time required to generate the scheme is only 1 / 5000 of the traditional functional knowledge integration method based on the DFS algorithm.

[0064] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A functional knowledge integration (FKI) method for innovative design based on deep reinforcement learning, comprising the following steps: 1) Constructing a functional knowledge graph (FKG) by matching the input and output of each discrete functional knowledge unit; 2) Modeling the core task of FKI, i.e., functional unit chain generation, as a Markov decision process (MDP) and solving it using the deep Monte Carlo algorithm in reinforcement learning; specifically including: 2.1) Represent the target function as two virtual nodes in FKG and , the former contains only output attributes and the latter contains only input attributes; then, the FKI task is defined as finding a fully matched path from to in FKG, i.e. the sum of the redundancies between functional units in the path is 0; further, by using the idea of dynamic programming, the complex FKI task is transformed into searching a partially matched path from to each time, i.e. a chain of functional units, and then iteratively completing the matching of all input / output by taking the unmatched input / output as a new target function and searching its chain of functional units. 2.2) Defining the sum of the number of redundancies in the path as the evaluation index of the functional unit chain, called global redundancy; optimizing this index can ensure the functional completeness of the obtained scheme and control the number of iterations required by the algorithm within a reasonable range; 2.3) Modeling the FKI task as an MDP, where each time step corresponds to the selection of a successor node in the path search process; 2.4) Construct a value network Q-network to predict the Q-value of each successor node from the target input node The network contains two feature extraction modules of convolution and LSTM, and a MLP prediction module. It receives seven pieces of information as input, including the current node, adjacent nodes, the number of redundancies between the current node and its successor nodes, the behavior trajectory, the global redundancy number, and the target functional requirements. It outputs the expected return value Q-value of each successor node. 2.5) Defining four reward functions, i.e., success, failure, process, and wall, to provide feedback to the agent in four scenarios: successfully reaching the target output node, reaching the maximum search length without reaching the final node, being in an intermediate node, and being in an intermediate node with no successor node. The failure reward is a non-zero reward based on the health-oriented hypothesis, allowing the agent to learn how to search healthily even when it fails. This reward can significantly improve the convergence speed and performance of the algorithm; 2.6) Training the value network using the Every-visit Monte-Carlo algorithm; defining the target policy as a greedy policy and the behavior policy as an epsilon-greedy policy improved based on the epsilon-greedy policy, with the epsilon value decreasing during the training process; 3) Merging the functional unit chains generated by the reinforcement learning algorithm into the final design scheme through an automatic or semi-automatic iterative framework; achieving an efficient FKI framework that takes the user's structured semantic description of the target function as input and outputs the corresponding functional design scheme.

2. The function knowledge integration method for innovation design based on deep reinforcement learning according to claim 1, wherein: Step 1) specifically includes: 1.1) Build a functional knowledge graph (FKG) with the minimum reuse unit of functional design knowledge - functional unit, and represent it as a directed graph ; Where each node represents a functional unit, and each edge represents the matching relationship between and ; The knowledge in the knowledge graph exists in the format of triplets ( , , ), which specifically represents that the output of the functional unit is matched by the input of the functional unit , and they are called predecessors and successors respectively; 1.2) The ontology of a functional unit is defined as having the following seven attributes: ID, Input, Output, Prior, Next, Category, and Carrier; where ID is the unique identifier of the functional unit; Input and Output represent the input and output of the functional unit, describing the basic content of the product function; Prior and Next reflect the position (or adjacency) relationship of the functional unit in the FKG, corresponding to the predecessor and successor, respectively; Category refers to the functional category carried by the functional unit, including conversion, support, storage, and excitation functions; Carrier represents the physical or technical prototype of the functional unit in reality; 1.3) The representation of input and output is the focus of functional representation, adopting a keyword plus parameter representation method; specifically, each keyword exists in the format of "modifier + core word"; each parameter exists in the format of "name + value range + unit"; each input or output can be composed of multiple keywords and parameters, and each keyword can have multiple modifiers but only one core word. 1.4) To save representation space and solve the problem of expression ambiguity, a bucket mechanism is designed for polysemy management; specifically, each word is divided into different buckets according to its word meaning, and is uniformly represented by the tag of the bucket; at this time, for any polysemy appearing in the functional unit entry process, the user decides its specific meaning by assigning it to different buckets; 1.5) The matching relationship between functional units is defined as follows: for functional units and , if any output is matched by any input, then it is said that functional units and are matched; and for a pair of input and output, only if their core words are in the same bucket, and the modifier and feature of the output contain those of the input, especially, if a parameter appears in both input and output, then its value domain in the input should be contained in that of the output, in order to be called matched; the number of unmatched input / output between adjacent functional units is taken as a quantitative indicator of the matching relationship, which exists in the form of a label on each edge of the functional knowledge graph, called redundancy number.

3. The function knowledge integration method for innovation design based on deep reinforcement learning according to claim 1, wherein: Step 3) specifically includes: 3.1) The structured description of the target function input by the user is taken as the input of the reinforcement learning algorithm in step 2), and a corresponding optimal functional unit chain is output; based on the fact that the FKI process has been converted from search to Markov decision process at this time, the functional unit chain can be generated at almost real-time speed in this step; 3.2) The unmatched input / output in the functional unit generated in step 3.1) is taken as the output / input of the target function in the next round of iteration, respectively; in this step, the user is allowed to adjust the existing unmatched input / output, i.e. switching to a semi-automatic iterative framework, and the allowed adjustments include but are not limited to deletion, addition, manual matching, parameter modification; 3.3) Steps 3.1) and 3.2) are iterated until all inputs / outputs in the path have been completely matched or confirmed by the user as completely matched.

Images