摘要
In reinforcement learning of long-term tasks, learning efficiency may deteriorate when an agent’s probabilistic actions cause too many mistakes before task learning reaches its goal. The new type of state we propose 8211; fixed mode 8211; to which a normal state shifts if it has already received sufficient reward 8211; chooses an action based on a greedy strategy, eliminating randomness of action selection and increasing efficiency. We start by proposing the combining of an algorithm with penalty avoiding rational policy making and online profit sharing with fixed mode states. We then discuss the target system and learning-controller design. In simulation, the learning task involves stabilizing of biped walking by using the learning controller to modify a robot’s waist trajectory. We then discuss simulation results and the effectiveness of our proposal.