| 研究課題/領域番号 |
19F19380
|
|---|
| 研究機関 | 中央大学 |
|---|
研究代表者 |
新妻 実保子 中央大学, 理工学部, 准教授 (10548118)
|
|---|
| 研究分担者 |
VINCZE DAVID 中央大学, 理工学部, 外国人特別研究員
|
|---|
| 研究期間 (年度) |
2019-11-08 – 2022-03-31
|
| キーワード | rule-base reduction / antecedent redundancy / fuzzy rule interpolation / fuzzy rule clustering / knowledge extraction / fuzzy control |
|---|
| 研究実績の概要 |
In this time period, according to the research plan, fuzzy rule-base reduction and fuzzy rule simplification possibilities were investigated. The fuzzy systems under investigation are using Fuzzy Rule Interpolation (FRI), which allows a fuzzy rule-base to be intentionally incomplete, while still being to correctly operate the (control) system. Automatically finding those rules which can be omitted is not a trivial task, but have been proved previously that it is possible to some extent.
A new method was developed and implemented which is capable of finding redundant parts (antecedents) in given fuzzy rules. This novel method works by checking the antecedents one by one of a given fuzzy rule and determines whether the antecedents in a fuzzy rule are strictly required or not. This new antecedent redundancy exploration method was implemented in the framework of the FRIQ-learning reinforcement learning method. This framework includes standard benchmark applications, hence the new antecedent redundancy investigating method can be compared to previous methods. The results are promising as this new method can create rule-bases less than half the size or even smaller than the previous reduction methods with the same benchmarks.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
The progress of the research is according to the submitted plans right now. New rule-base reduction methods for the FRIQ-learning have been successfully designed, implemented and evaluated. Two papers have been prepared and accepted for publication at high ranking conferences in the field. A third paper is currently in the works. An interface has been developed and is under implementation between the FRIQ-learning reinforcement learning framework and the fuzzy automaton-based complex behaviour simulation system (Strange Situation Test simulation).
|
| 今後の研究の推進方策 |
The research can be continued as planned. It would be worth investigating the combination of the two newly developed fuzzy rule-base and fuzzy rule simplification methods (antecedent redundancy exploitation and clustering-based reduction). This could yield a more efficient method for simplifying the knowledge base of behaviour based systems like the proposed football simulation and the SST model. The next milestone is to efficiently interconnect the FRIQ-learning and the SST model, and to construct appropriate reward functions for the reinforcement learning process for parts of the SST model.
|
