Towards a framework for reinforcement learning with artificial neural networks

Reinforcement learning is a quite general learning paradigm that can be used to solve a large set of problems. For complex problems it has been shown that by using task decomposition it may be possible for the system to learn faster. One common approach is to construct systems with multiple modules, where each module learns a sub-task. We present a parallel learning method for agents with an actor–critic architecture based on artificial neural networks. The agents have multiple modules, where the modules can learn in parallel to further increase learning speed. Each module solves a sub-problem and receives its own separate reward signal with all modules trained concurrently. We use the method on a grid world navigation task and show that parallel learning can significantly reduce learning time.