Reinforcement Learning Agent for Knowledge Graph Relation Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual skill matching in industrial systems, such as manufacturing, is time-consuming, expensive, and prone to human errors and cognitive biases, especially when dealing with large amounts of incomplete knowledge graphs.
Innovation Solution
A computer-implemented method using a reinforcement learning agent for policy-guided transitions in a knowledge graph, combining vector space embeddings and recurrent neural networks, with supervised training and logical rules to predict and rank relations, enabling automatic completion of missing links.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual skill matching is performed by domain experts, then reliability of predictions is improved, but productivity is worsened due to time-consuming and expensive manual processes
Solution Approach 1:
The patent replaces the manual mechanical process of expert inspection with an automated computational system using reinforcement learning agents and knowledge graphs. The system automatically performs skill matching by learning from existing data patterns, eliminating the need for manual domain expert intervention while maintaining prediction reliability through algorithmic consistency and scalability.
Solution Approach 2:
The system enables self-service by allowing the knowledge graph and reinforcement learning model to automatically infer missing relations and perform skill matching without external expert input. The model learns from existing data and autonomously completes the matching process, making the system self-sufficient for routine operations.
2Measurement precision
If manual inspection of knowledge graphs is performed, then measurement precision of missing links is improved, but loss of time is worsened due to extensive manual review required
Solution Approach 1:
The patent replaces manual inspection with automated reinforcement learning agents that systematically analyze the knowledge graph structure and infer missing links. The computational approach maintains measurement precision by learning from patterns in existing data while dramatically reducing the time required through parallel processing and algorithmic efficiency.
Solution Approach 2:
The system performs preliminary action by pre-training the reinforcement learning model on existing knowledge graph data before actual skill matching tasks. This pre-learning phase enables the model to quickly infer missing relations during operation, reducing real-time prediction time while maintaining accuracy through previously learned patterns.
3Manufacturing precision
If manual skill matching is used, then manufacturing precision of skill- operation matching is improved, but loss of time is worsened due to extensive expert review
Solution Approach 1:
The patent replaces manual expert matching with automated reinforcement learning agents that learn optimal skill-operation pairings from existing data. The system achieves manufacturing precision through consistent application of learned patterns and can process multiple matching tasks simultaneously, dramatically reducing the time required compared to sequential manual review.
Solution Approach 2:
The system enables continuous operation by automatically processing skill matching requests without interruption by manual review processes. The reinforcement learning model continuously learns from new data and can handle multiple matching tasks in sequence or parallel, eliminating the discontinuities and delays inherent in manual expert processes.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a computer-implemented method for prediction of relations in a knowledge graph, KG, representing an industrial system and stored in a database, the method comprising the steps of receiving (S1) via an interface a query, Q, comprising a subject-predicate pair (s, p) corresponding to a subject entity of the industrial system represented by a source node in the knowledge graph, KG, and to a relation of said subject entity represented by an edge from said source node; performing (S2) automatically policy-guided transitions by a reinforcement learning agent starting from the source node of the knowledge graph, KG, along edges of the knowledge graph, KG, to target entities represented by target nodes of the knowledge graph, KG, to extract different paths between the source node and the target nodes, wherein for each performed transition, a transition probability is obtained from a discrete stochastic policy; calculating (S3) a path probability for each extracted path on the basis of the obtained transition probabilities of the transitions between the source node and the target node of the respective path; and ranking (S4) the extracted paths according to the calculated path probabilities to select a target entity of said industrial system output via said interface.