Name of Paper

Shared experience learning on a pair of autonomous mobile robots


Authors

I.D. Kelly and D.A. Keating


Published

Distributed Autonomous Robotic Systems 2, 1996, Springer-Verlag, p. 434.


Abstract

Previously we have developed an algorithm, for mobile robots, to learn how to reactively avoid obstacles based on sets of fuzzy automata. This paper deals with two robots learning simultaneously using the same strategy, but with the experiences of one robot being passed to the second robot. To this end two small autonomous mobile robots have been constructed which are equipped with a radio communication system and an ultrasonic sonar for detecting obstacles. The obstacle detection system consists of three sets of ultrasonic sonar transducers, one set looking forward, another to the front-left and the third to the front-right.

The learning algorithm is based on fuzzy automata. With two bi-directional motors there are nine different possible actions, if the only motor actions are forwards, backwards and stop. We have defined five different circumstances that the robot can find its self in: no object near robot, obstacle in distance to the left, obstacle in distance to the right, obstacle near the right and obstacle near the left. Each of these different circumstances has its own fuzzy automaton associated with it. Each automaton, which is effectively a set of motor actions, has a set of probabilities of taking the associated action. A weighted roulette wheel technique is used to randomly select the most appropriate action for the given input (obstacle position). The action with the highest probability is the most likely to be chosen. The chosen action is executed for a short period of time and is then evaluated. If the action was successful its probability of being selected is increased. If the action was a failure its probability of being selected is decreased. The rules used to evaluate the robots performance are that if there is no object within range it is good to go forward but if an object is relatively near it is good to get further away from it. These rules are general and hence will not give the robot any information about what actions to take. The information transmitted from one robot to the other consists of: the circumstance, the chosen action and the resulting fitness.

The results clearly show that shared experience learning is faster and more robust than individual learning. Shared experience learning also allows increased learning rates to be exploited which the effects of noise prevents in individual learning. The authors are currently investigating the effects of true mutual learning where two or more robots share each others experiences.