COMBINING REINFORCEMENT LEARNING AND GENETIC ALGORITHMS TO LEARN BEHAVIOURS IN MOBILE ROBOTICS

摘要

Reinforcement learning is an extremely useful paradigm which is able to solve problems in those domains where it is difficult to get a set of examples of how the system should work. Nevertheless, there are important problems associated with this paradigm which make the learning process more unstable and its convergence slower. In our case, to overcome one of the main problems (exploration versus exploitation trade off), we propose a combination of reinforcement learning with genetic algorithms, where both paradigms influence each other in such a way that the drawbacks of each paradigm are balanced with the benefits of the other. The application of our proposal to solve a problem in mobile robotics shows its usefulness and high performance, as it is able to find a stable solution in a short period of time. The usefulness of our approach is highlighted through the application of the system learnt through our proposal to control the real robot.