State and action space construction using vision information

To apply reinforcement learning to the real world, it needs pre-processed sensor data which is adequate for action learning. Since it is difficult to construct state space and learn an appropriate action simultaneously, we assume that an estimation is given to each step of action, whether it is good or bad. Under this condition, we propose a method of dividing and clustering the state space. The TRN (topology representing network) is a vector quantization algorithm, and it can preserve topology in the input space. We apply the TRN algorithm to our problem with dynamically increasing nodes and the idea of a radial basis function.