Application Of Radial Basic Function Networks And Instance Based Learning In Reinforcement Learning

An important goal of Artificial Intelligence(AI) is to design an agent that can complete a given task in complex environment. Machine learning is an important component of AI .Reinforcement Learning(RL) is a new branch of machine learning and because of its feature it is proper to be used to design an agent that learns from experience. To RL, the main idea of choosing a policy bases on reward signals that come from environment, i.e. an agent can obtain knowledge while it interacts with environment continually. Unlike that supervised learning needs input-output pairs it only uses reward signals to improve its behaviors. Recently, there are more works focused on RL. It is proposed as an important technology to design an agent. Reinforcement Learning appeals to many researchers because of its generality. In RL, an agent is simply given a goal to achieve, then the agent learns how to achieve that goal by trial-and-error interactions with its environment. Traditional RL algorithms are restrained in limited discrete space and use tabular to represent the value function .However, most real systems' input space which we meet with are continuous so that it is impossible to use tabular. A good alternative is to use a approximator while it is not simply replaces tabular with a approximator because of disconvergence. However, some study shows that it is appropriate to use local approximator to represent value function. This thesis mainly utilizes local approximation to represent value function. The main work is as follows:(1) proposes to use RBF to generalize value function;(2) makes use of Instance-Based Learner as value function approximator to improve system performance;(3) Experiments have been done to show some good results.