Paper: An Attempt to Generalize AI - Part 4: Modeling Efficiency

Abstract: This is the fourth in a series of articles attempting to give an overview of how minds may work and how similar systems could be implemented in computers. The first article described a probabilistic, hierarchical modeling system, based on patterns, which are sets of pattern instances, intended to provide a general ontology. The second article described the use of this for planning actions. A serious issue with AI systems is ensuring that the computation that is done is useful. A system like this finds patterns, then patterns based on the patterns, and so on. Only a small amount of this computation will be relevant. Unless something is done to prevent it, there will be a huge proliferation of pattern instances that are of little use in the making of predictions. This article explores this issue, and starts to consider, very broadly, what may be involved in reducing the number of pattern instances in the model. This will be built on in later articles, which will discuss specific methods of achieving what is discussed here.

Read entire paper here (PDF).