next up previous contents
Next: Conclusions and Contributions Up: Limitations Previous: Well-Behaved, Smooth Representations are

No High Level Goals, Motivations, or Sequences of Actions

The Action-Reaction Learning framework is clearly only recovering lower-order reactionary behaviour. The actions and reactions are mapped in an almost direct manner and the short-term memory really only contains the last stimulus that triggered the system. Thus, higher order behaviours that involve complex chains of actions are not addressed. For instance, there are no mechanisms for dealing with sequences of actions that are strung together into some more complex meta-action. There is also no representation of the actions forming sub-goals and these forming then goals in a sequential evolution. The ARL system does not identify goals, plans or motivations which are some of the essential concepts in higher order behaviour. Therefore, we only addressed a basic definition of behaviour as a stimulus-response mapping. The transition to higher order behaviour is an interesting issue and remains a topic of future work.

Tony Jebara