A Framework for Aggregation of Multiple Reinforcement Learning Algorithms

Aggregation of multiple Reinforcement Learning (RL) algorithms is a new and effective technique to improve the quality of Sequential Decision Making (SDM). SDM is very common and important in various realistic applications, especially in automatic control problems. The quality of a SDM depends on (discounted) long-term rewards rather than the instant rewards. Due to delayed feedback, SDM tasks are much more difficult to handle than classification problems. Meanwhile, in many SDM tasks, the feedback about a decision is often in the form of evaluation rather than instruction. Therefore, supervised learning techniques are not suitable in these tasks. To tackle these difficulties, RL methods are investigated. Although many RL algorithms have been developed, none is consistently better than the others. In addition, the parameters of RL algorithms significantly influence learning performances. Successful RL applications depend on suitable learning algorithms and elaborately selected learning parameters, but there is no universal rule to guide the choice of algorithms and the setting of parameters. To handle this difficulty, a new multiple RL system - the Aggregated Multiple Reinforcement Learning System (AMRLS) is developed. In this proposed system, each RL algorithm (learner) learns individually in a learning module and provides its output to an intelligent aggregation module. The aggregation module dynamically aggregates these outputs by using some intelligent aggregation methods and provides a decision of action. Then, all learners take the action and update their policies individually. The two processes are performed alternatively in each learning episode. Because of the intelligent and dynamic aggregation, AMRLS has the ability to deal with dynamic learning problems without the need to search for the optimal learning algorithm or the optimal values of learning parameters. It is claimed that several complementary learning algorithms can be integrated in the AMRLS to improve the learning performance in terms of success rate, robustness, confidence, redundance, and complementariness. There are two strategies for learning an optimal policy by using RL methods. One is based on the Value Function Learning (VFL) strategy, which learns an optimal policy expressed as a value function. The Temporal Difference (TD) methods are examples of this strategy and they are called TDRL in this dissertation. The other strategy is based on the Direct Policy Search (DPS), which directly searches for the optimal policy in the potential policy space. The Genetic Algorithms (GAs)-based search algorithms are instances of this strategy and they are named GARL. Both of the strategies exhibit advantages and disadvantages. A hybrid learning architecture of GARL and TDRL, HGATDRL, is proposed to combine them together. HGATDRL uses an off-line GARL approach to learn an initial policy first, and then updates the policy on-line by using a TDRL approach. This new learning method enhances the learning ability of RL learners in AMRLS. The AMRLS framework and HGATDRL method are tested on several SDM problems, including the maze world problem, pursuit domain problem, cart-pole balancing system, mountain car problem, and flight control system. The experimental results show that the proposed framework and method can enhance the learning ability and improve learning performance of a multiple RL system.