Learning ecient reactive behavioral sequences from basic re in a goal-oriented autonomous robot

This paper describes a reinforcement connec-tionist learning mechanism that allows a goal-directed autonomous mobile robot to adapt to an unknown indoor environment in a few trials. As a result, the robot learns e±cient reactive behavioral sequences. In addition to quick con-vergence, the learning mechanism satis¯es two further requirements. First, the robot improves its performance incrementally and permanently. Second, the robot is operational from the very beginning, what reduces the risk of catastrophic failures (collisions). The learning mechanism is based on three main ideas. The ¯rst idea applies when the neural network does not react properly to the current situation: a ¯xed set of basic re-°exes suggests where to search for a suitable ac-tion. The second is to use a resource-allocating procedure to build automatically a modular net-work with a suitable structure and size. Each module codi¯es a similar set of reaction rules. The third idea consists on concentrating the ex-ploration of the action space around the best ac-tions currently known. The paper also reports experimental results obtained with a real mobile robot that demonstrate the feasibility of our ap-proach.