Unoriented Reinforcement Learning by Memory Repetition Model and its Verification by Language-related Simulation

Abstract

A brain mechanism hypothesis called a “punisher” defined as activating a large number of nodes is proposed, and under this hypothesis, a model of reinforcement learning is proposed, whose purposes are not given, but rather the model decides its purposes to avoid the failures expected in the future. The model is a recurrent neural network (RNN), and is characterized as follows: 1) the important concepts are captured by reinforced (or repeated) “autographical memory”, i.e., the series of active nodes and 2) the model acquires knowledge and decides the action using the simple mechanism that any repetition is reduced when a punisher is given. It is also shown that an autobiographical memory is actually created in a small-scale RNN. It is also shown the simulations that verify whether a given language-related task could be performed according to the rules of the model.