5 Criteria an agent might use in the search for a good internal model.

5.4 Complexity

The question naturally arises, why isn't the ideal model simply composed of past data. It would have zero error, minimum possible volume and minimal cost. Our experience tells us that this is not generally a good model at prediction but, equally, our experience does not tell us how to formalise this in our model of learning by agents.

One way to bias the agent towards agent models that are more likely to be productive is to limit its agent modelling language to rule out models comprising nothing more than past observations. This, however, would probably mean a more restricted language than is sometimes desirable and would probably not rule out all analogous situations within the new language.

A second way is to include some element of complexity to guide model search, i.e. in certain circumstances (e.g. when error rates are almost equal) bias the agent to choose the simpler model. We are not claiming that the simpler model is a priori more likely to be correct (Quine [23], Pearl [21]), just that this is an effective heuristic for search within such open-ended spaces of language expressions.

There are many possible ways to measure complexity (Edmonds [6]) and each will result in a slightly different search pattern. All we require is that the measure be practically computable and that it can act as a limit to naive depth-first strategies that might be applied by an agent.

Modelling Learning as Modelling - 23 FEB 98

[Next] [Previous] [Up] [Top] [Contents]

Generated with CERN WebMaker