Decision Boundary Estimation Using Reinforcement Learning for Complex Classification Problems

In this paper, we propose a method for quickly training a binary support vector machine (SVM) classifier for recognizing valid input spaces in high-dimensional, highly constrained systems by using reinforcement learning to find inputs along the decision boundary of the classifier while minimizing the number of runs of a simulation representing the system. We find training points by first defining an optimization problem where the action space consists of points to test, and the reward is a function that searches for points that are close to violating the given constraints and are a sufficient distance from one another. After formulating this process, we use a Q-learning framework to find inputs that maximize the reward, and then use these inputs to train the classifier so the decision boundary is quickly well-defined. The efficacy of this approach is then shown in simulations.