Abstract
We consider the problem of designing local reinforcement learning rules for artificial neural network (ANN) controllers. Motivated by the universal approximation properties of ANNs, we adopt an ANN representation for the learning rules, which are optimized using evolutionary algorithms. We evaluate the ANN rules in partially observable versions of four tasks: the mountain car, the acrobot, the cart pole balancing, and the nonstationary mountain car. For testing whether such evolved ANN-based learning rules perform satisfactorily, we compare their performance with the performance of SARSA() with tile coding, when the latter is provided with either full or partial state information. The comparison shows that the evolved rules perform much better than SARSA() with partial state information and are comparable to the one with full state information, while in the case of the nonstationary environment, the evolved rule is much more adaptive. It is therefore clear that the proposed approach can be particularly effective in both partially observable and nonstationary environments. Moreover, it could potentially be utilized toward creating more general rules that can be applied in multiple domains and transfer learning scenarios.