A connectionist system approach for learning logic programs

In this paper, we show that temporal logic can be learnt effectively by a connectionist system. In contrast with other connectionist approaches in this context, we focus more on learning rather than knowledge representation. In order to learn from temporal logic values, the paper proposes a general three-layer connectionist system regardless of the number of logic rules, a condition which must have been satisfied in previous approaches. A mapping function is proposed to convert logic rules to the proper connectionist system´s inputs. Then a simulation study is carried out for muddy children puzzle. The results of the study suggest that an agent embedded with a connectionist system can learn temporal logic efficiently. It is observed that the connectionist system can increase its performance and make fewer mistakes while encountering with more produced cases of given logical rules.