Abstract
This paper proposes two kinds of complex-valued Profit Sharing algorithm for learning in environments involving perceptual aliasings. These algorithms are implementations in complex-valued reinforcement learning framework. The CVRL framework aims to solve perceptual aliasing problem by using context. The CVRL is characterized by complex-valued action values and a internal reference value. The internal reference value is used to distinguish confused states and revised at each step. As another implementation in the framework, Q-learning has been proposed. Several experimental results have demonstrated the effectiveness of the Q-learning. However, Q-learning suffers from a disadvantage in parameter settings because of many parameters. Therefore, this paper focuses on Profit Sharing which has less parameter than Q-learning and proposes complex-valued Profit Sharing. Furthermore, this paper also describes a technique named multiple phase changes. This technique enables the agent to revise the internal reference value adaptively and realizes more sufficient learning in environments which have confused states at various intervals. Simulation results support that proposed method works well in such environments.
