Receding Horizon Cache and Extreme Learning Machine based Reinforcement Learning
Path planning is a difficult part of the navigation task for the mobile robot under dynamic and unknown environment. It needs to solve a mapping relationship between the sensing space and the action space. The relationship can be achieved through different ways. But it is difficult to be expressed by an accurate equation. This paper uses multi-layer feedforward artificial neural network (ANN) to construct a path-planning controller by its powerful nonlinear functional approximation. Then the path planning task is simplified to a classified problem which are five state-action mapping relationship. One reinforcement learning method, Q-learning, is used to collect training samples for the ANN controller. At last the trained controller runs in the simulation environment and retrains itself furthermore combining the reinforcement signal during the interaction with the environment. Strategy based on the Combination of ANN and Q-learning is better than using only one of the two methods. The simulation result also shows that the strategy can find the optimal path than using Q-learning only.