A Framework for Generation and Management of Adaptive Rules Based on Reinforcement Learning

Abstract

The invention discloses a frame for generating and managing an adaptive rule based on reinforcement learning. According to the frame, a mapping relationship is established between an adaptive scene and a reinforcement learning algorithm. The frame comprises a designing period and an operating period. In the designing period, the reinforcement studying algorithm is used and adaptive rules are respectively generated for aiming at a plurality of representative target settings. In the operating period, reasoning based on cases is utilized for finding and activating the adaptive rule which is most suitable for the current target setting is found and activated, and furthermore the adaptive rule is continuously updated by means of the reinforcement learning algorithm. The frame has a technical effect of improving a traditional rule-based adaptive method and is advantageous in that 1, an automatic generating algorithm for the adaptive rule which can ensure matching between the generated rule and the target is supplied; and 2, an adaptive rule evolution algorithm which can make the rule set perform evolution that matches a target change after occurrence of the target change is supplied.

Description

Technical field: [0001] The invention relates to a framework for generating and managing adaptive rules based on reinforcement learning, which is suitable for supporting automatic generation and evolution of rules in an adaptive system. It belongs to the field of software technology. Background technique: [0002] An adaptive system refers to a system that can monitor changes in the environment and the system itself, and can dynamically modify its behavior and structure to continuously meet the goals. After sensing and analyzing the change, the adaptive system needs to plan in real time, that is, decide how to adjust the components or parameters of the software to continuously meet the goal. Among the common planning methods for adaptive systems, rule-based methods use rules to predefine adaptive strategies. Such methods usually maintain a set of CA (Condition-Action) rules, and each rule specifies that when a specific condition is true, the system should perform a specifi...

AI Technical Summary

Problems solved by technology

The rule-based planning method has the advantages of being readable, easy to modify, and efficient in the planning process, but this type of planning method also has its shortcomings: 1. Since the rule set is usually defined manually, it cannot be guaranteed that it will achieve the goal optimally; 2. , Since the defined rule set usually does not change anymore, it cannot adapt to the goals that may change dynamically

Images