Interpretable agent decision-making method, intelligent system and application of system

Abstract

The interpretable agent decision-making method comprises the following steps: extracting features in an application scene to establish a knowledge base and a knowledge base configuration file; performing symbolization expression on the knowledge base, and constructing a causal reasoning network; establishing connection between the factorization machine training layer and the causal reasoning network, and performing data training in a mode of growing new nodes in the causal reasoning network; and replacing a factorization machine layer with a single-layer full-connection layer, and then training. The process of explaining the decision of the agent comprises the following steps: tracing the decision, monitoring the node activation condition of the causal reasoning network, and monitoring a corresponding knowledge base and a knowledge base configuration file; and analyzing the weight matrix of the full connection layer, and finding out the characteristics influencing the decision. The monitoring and analysis results serve as traceability records and are output to interact with the user offline, and an intelligent model is established for understanding and trust of the user. According to the invention, the problem of uninterpretability of agent decision is solved, and training can be carried out to improve intelligence.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and particularly relates to an interpretable intelligent body decision-making method, an intelligent system and an application of the system. Background technique [0002] Most AI models in the field of artificial intelligence use deep models and reinforcement learning models or a combination of the two models, requiring big data training. These technologies can greatly improve the intelligence of the models, but the problem is that they are opaque and not interpretable , unable to explain their decisions to users. Although the expert system built on the knowledge base is interpretable, its performance is far from that of today's machine learning models and cannot be trained. SUMMARY OF THE INVENTION [0003] In order to overcome the above-mentioned shortcomings of the prior art, the purpose of the present invention is to provide an interpretable agent decision-making method, a...

AI Technical Summary

Problems solved by technology

[0002] Most AI models in the field of artificial intelligence use deep models and reinforcement learning models or a combination of the two models, which require big data training. These technologies can greatly improve the intelligence of the models, but the problem they bring is that they are opaque and not interpretable. , unable to explain their decisions to users

Although the expert system based on the knowledge base is interpretable, its performance is far from that of today's machine learning models and cannot be trained

Images