next up previous contents
Next: Visual Inputs and Outputs Up: Discussion and Properties Previous: Probabilistic and Bayesian Machine

Cognitive Concepts

The simplicity of Action-Reaction Learning is reminiscent of the behaviourists' model. However, the model also addresses some of the limitations and hurdles associated with the behaviourists' viewpoint. ARL also employs cognitive concepts that are more recent than behaviourism. Some cognitive concepts such as short term memory and attentional mechanisms are contained in the model. One example is an explicit attentional window with memory decay [16]. This is implemented to analyze past interactions in a short-term memory sense. The window also forms a coarse state model which allows the system to recall its own previous actions and maintain some self-consistent behaviour. Eventually, the approach could be made adaptive and more complete notions of internal state could be included (i.e. spanning variable length memory). The subsequent learning system would thus exhibit superior generalization and more realistic behaviour without extending the problem beyond the capacity of contemporary machine learning techniques.



Tony Jebara
1999-09-15