A performance comparison of TRACA - an incremental on-line learning algorithm

摘要

TRACA (Temporal Reinforcement-learning and Classification Architecture) is a learning system intended for robot-navigation tasks. One problem in this area is input-generalisation. Input generalisation requires learning a small set of internal states which represent useful abstractions of the much larger set of actual states. As such, the input-generalisation problem is fundamentally similar to the classical problems of classification, concept learning and discrimination. The priorities when evaluating a system for on-line robot learning include a small number of trials, predictive accuracy and minimal parameter tuning. Other requirements are the ability to learn without predefined classes (i.e. classes must be learned during training) and an efficient and adaptable representation. This paper evaluates the performance of TRACA, a new learning algorithm, on a number of common classification tasks. The same set of parameters is used to obtain all TRACA's results, which are then compared to the results obtained by other well-known algorithms. On most tasks, TRACA's predictive accuracy is within a few percent of the best performing systems compared. Furthermore, TRACA's result is often achieved with less training experience. In a final experiment TRACA is trialled on a robot navigation task that requires discrimination of a number of discrete locations.