Resumen de Regularized approximate policy iteration using kernel for on-line reinforcement learning
Gennaro Esposito
By using Reinforcement Learning (RL), an autonomous agent interacting with the environment can learn how to take adequate actions for every situation in order to optimally achieve its own goal. RL provides a general methodology able to solve uncertain and complex decision problems which may be present in many real-world applications. RL problems are usually modeled as a Markov Decision Processes (MDPs) deeply studied in the literature. The main peculiarity of a RL algorithm is that the RL agent is assumed to learn the optimal policies from its experiences without knowing the parameters of the MDP. The key element in solving the MDP is learning a value function which gives the expectation of total reward an agent might expect at its current state taking a given action. This value function allows to obtain the optimal policy. In this thesis we study the capacity of SVR using kernel methods to adapt and solve complex RL problems in large or continuous state space. SVR can be studied using a geometrical interpretation in terms of optimal margin or can be seen as a regularization problem given in a Reproducing Kernel Hilbert Space (RKHS) SVR have good properties over the generalization ability and as they are based a on convex optimization problem, they do not suffer from sub-optimality. SVR are non-parametric showing the ability to automatically adapt to the complexity of the problem. Accordingly, applying SVR to approximate value functions sounds to be a good approach. SVR can be solved both in batch mode when the whole set of training sample are at disposal of the learning agents or incrementally which enables the addition or removal of training samples very effectively. Incremental SVR finds the appropriate KKT conditions for new or updated data by modifying their influences into the regression function maintaining consistence in the KKT conditions for the rest of data used for learning. In RL problems an incremental SVR should be able to approximate the action value function leading to the optimal policy. Accordingly, computation load should be lower, learning speed faster and generalization more effective than other existing method The overall contribution coming from of our work is to develop, formalize, implement and study a new RL technique for generalization in discrete and continuous state spaces with finite actions. Our method uses the Approximate Policy Iteration (API) framework with the BRM criterion which allows to represent the action value function using SVR. This approach for RL is the first one we know using SVR compatible to the agent interaction- with-the-environment framework of RL which shows his power by solving a large number of benchmark problems, including very difficult ones, like the bicycle driving and riding control problem. In addition, unlike most RL approaches to generalization, we develop a proof finding theoretical bounds for the convergence of the method to the optimal solution under given conditions.
