File(s) under permanent embargo
Reinforcement learning using continuous states and interactive feedback
conference contributionposted on 01.01.2019, 00:00 authored by A Ayala, C Henríquez, F Cruz Naranjo
Research in intelligent systems field has led to different learning methods for machines to acquire knowledge, among them, reinforcement learning (RL). Given the problem of the time required to learn how to develop a problem, using RL this work tackles the interactive reinforcement learning (IRL) approach as a way of solution for the training of agents. Furthermore, this work also addresses the problem of continuous representations along with the interactive approach. In this regards, we have performed experiments with simulated environments using different representations in the state vector in order to show the efficiency of this approach under a certain probability of interaction. The obtained results in the simulated environments show a faster learning convergence when using continuous states and interactive feedback in comparison to discrete and autonomous reinforcement learning respectively.