Proposal of an Exploitation-oriented Learning Method on Multiple Rewards and Penalties Environments and the Design Guideline

Among machine-learning approaches, reinforcement learning (RL) focuses most on goal-directed learning from interaction. Despite important applications, RL is difficult to design to fit real-world problems because, first, interaction requires too many trial-and-error searches and, second, no guidelines exist on how to design reward and penalty signal values. We are interested in approaches treating reward and penalty signals independently and not assigning them values. We also want to reduce the number of trial-and-error searches by strongly enhancing successful experience — a process known as exploitationoriented learning (XoL). Though there are many XoL methods, they cannot apply to multiple rewards and penalties environments adequately. In this paper, we propose a new XoL method that can treat multiple rewards and penalties effectively. We present simulation and experimental results to show the effectiveness of our proposal. Furthermore, we describe the design guideline about rewards and penalties for the XoL methods.