Concatenated Dynamic Reinforcement Learning for Multi-staged Tasks (MST)

There exists a class of complex tasks that can be distinctly separated into multiple concatenated simpler tasks [1-2]. This class of complex and continuous motion processes can then be considered as multiple motion sub-processes, and a switch that brings these sub-processes together in accordance to an input function. This paper validates the potential value of this approach by proposing a method to train a complex task through a series of sub-tasks, and then concatenating them through a switch. Through the classic simulation physical task of OpenAI Gym[3]: cartpole, the paper demonstrates the viability and value of this approach. An example complex task of moving the pendulum back and forth between to points was used. A set of reward and switching functions were developed to achieve the complex task though the use of Deep Q-network (DQN) [4]. Results show that the proposed approach achieves the goal within a limited number of training cycles that otherwise cannot be met by other traditional means. Alternative reinforcement learning methods would result in more input parameters, larger deep learning networks, and larger accumulative training cycles.