A Design Method of Fuzzy Logic Controller by Using Q Learning Algorithm
Xin Zhou, Cai Zhi Fan, Jun Wu · 2018
As one of Reinforcement Learning (RL), Q learning algorithm has been applied in many fields of dynamic programming, and its Q Neural Network (Q-NN) can map the states of the environment to the corresponding control actions. Q learning algorithm can reach or exceed human level in some video games. Nevertheless, the function of a neural network is similar to a black box. When the network fails in some control tasks, the adjustment of the network by the people is very difficult to achieve. In contrast, due to the semantic control rules of the Fuzzy Logic Controller (FLC), incorrect rules can be artificially adjusted, however, a completely manual control rules design is a burdensome task. Therefore, in this paper, we propose a new method for the FLC to learn rules from well-trained neural networks and then fine-tune incorrect rules based on human knowledge. Finally, we experimented with navigational task and showed that the FLC we designed is able to accomplish the task, which achieved all of the reaching-target goals in 100,000 tests. (all the experiments results and relative codes are available on https://github.com/inksci/logic-controller).