リザーバコンピューティングによる目的切替可能な強化学習

抄録

This paper proposes a multi-objective switchable motor control by reinforcement learning with a reservoir computing. In general, the reinforcement learning control for respective given objectives is achieved by unique function approximators, which are often represented by deep neural networks. This fact means that, to switch the control to the one for the other objective, the function approximator should be switched to the one corresponding to the given objective. Storing the function approximators corresponding to all objectives in a memory or a storage, however, is wasteful and not practical. Hence, an adaptive function approximator, which can switch the objective to be represented only by changing inputs, is proposed. This adaptive function approximator is given by the reservoir computing. As a result, a cart-pole system acquired the controls for two switchable objectives: height and angular velocity of a pole.