www.bccn-goettingen.de

Temporal sequence learning requires at least two inputs that follow each other in time. Initially the learner will respond only to the later stimulus, while after learning it ought to be able to react to the earlier one. This can be achieved by a variety of learning rules, which will be compared in this article. Two main classes will be discussed: reinforcement learning and correlation based rules. When comparing these two classes we will mainly focus on their conditions of convergence. First we will discuss the rules in an open-loop condition, where the output of the learner will not lead to any behaviour. Here we find that both classes have rather similar properties. Differences become more pronounced in the closed-loop case, where we observe that the different convergence conditions of both classes lead to different characteristics. It appears that reinforcement learning is better suited to learn goal-directed behaviour, while correlation based learning should be employed when trying to learn maintaining homoeostasis.