查看文章

Q-learning is a Reinforcement Learning method where the learner builds incrementally the Q-function which estimates the future rewards for taking actions from a given state. This paper studies the application of Q-learning on Process Control problems, more precisely on Neutralization Processes. As the process to be studied is non-linear, 9 states are selected. Each state has 5 possible actions (unless in goal state, which has 3 actions), that corresponds to variations of the control signal. Softmax and ε-greedy policies are applied to select the actions on a laboratory pH plant. On-line results show that the controllers are able to learn how to control adequately the process.