We have tried out Q-learning in a couple of different configurations, with a limited amount of success in training our robot. The main problem with Q-learning is that we have a very large number of possible states, or positions, that the robot arm can be in. This means that gaining a lot of knowledge about any one position by repeated trials is very difficult. We are going to introduce a different approach using genetic algorithms to generate our movement actions.
Moving the robot arm requires coordination of three motors simultaneously to create a smooth movement. We need a mechanism to create different combinations of motor movement for the robot to test. We could just use random numbers, but that is inefficient, and ...
