Adaptable neuro production systems

Abstract

Connectionist production systems are neural network realizations of production rule-based systems. The connections are adjusted to a given set of rules to allow the system to perform reasoning. Adaptable connectionist production systems are introduced in this paper. They allow adaptation of the already pre-calculated connections to new data. The production rules are used to initialize the connection weights after which training with data occurs. At any time of the neural network operation, a set of updated rules can be extracted as a current knowledge base accumulated by the network. Using a set of rules for initializing a connectionist architecture before training may result in: (1) increase in the speed of training; (2) increase in the robustness of the neural network against the ‘catastrophic forgetting’ phenomenon; (3) better explanation of the learned by the network knowledge from data. In general, the proposed method facilitates building flexible and adaptable neuro-fuzzy production systems. This is demonstrated on a case problem of chaotic time series prediction.