Hyo-Byung Jun

Performance improvement of evolution strategies using reinforcement learning

We propose a new type of evolution strategies combined with reinforcement learning. We use the change of fitness occurred by mutation to form the reinforcement signals which estimate and control the step length of mutation. With this proposed method, the convergence rate is improved. Also, we use Cauchy distributed mutation to increase the global convergence faculty. Cauchy distributed mutation is more likely to escape from a local minimum or move away from a plateau than Gaussian distributed mutation. After an outline of the history of evolution strategies, we explain the evolution strategies combined with the reinforcement learning, that is reinforcement evolution strategies. Performance of the proposed method is estimated by comparison with conventional evolution strategies on several test problems.

Fuzzy inference-based reinforcement learning of dynamic recurrent neural networks

This paper presents the fuzzy inference-based reinforcement learning algorithm of dynamic recurrent neural network, similar to the psychological learning scheme of the higher animals. The proposed method follows the way linguistic and conceptional expressions have an effect on humans behavior by reasoning reinforcement based on fuzzy rules. The intervals of fuzzy membership functions are found optimally by genetic algorithms. By using the recurrent neural network composed of dynamic neurons as action-generation network, not only the current state but also the past state is considered to make an action in dynamical environment. We show the validity of the proposed learning algorithm by applying it to the inverted pendulum control problem.

Behavior learning and evolution of collective autonomous mobile robots based on reinforcement learning and distributed genetic algorithms

In this paper, we present the reinforcement learning and distributed genetic algorithm based behavior learning of the distributed autonomous mobile robots. The internal reinforcement signal for the reinforcement learning is generated by fuzzy inference, and dynamic recurrent neural networks are used as action generation module. We adopt the distributed genetic algorithms for the cooperative behavior emergence. We show the validity of the proposed learning and evolution algorithm by computer simulation.