Reinforcement learning algorithm based on immune tolerance

A novel reinforcement learning algorithm, which employs an intelligent mechanism called artificial immune tolerance process, is proposed for problems of the trap into a local extremum and the divergence of the state values when introducing the value function approximation in reinforcement learning, and also for reducing the learning factor´s influence on the whole processing. This algorithm adjusts the weights with TD(λ) from the function of immune tolerance. During the learning process, weights control the state values based on function approximation. When the error is greater than a certain threshold, the system uses immune tolerance to optimize the weights. Otherwise, the system selects the best strategy according to the situation. After performance analyses and simulations, the results show that the new algorithm can have smaller errors, faster global search, greater diversity, and less influence of the learning factors.