Planning As Modelling in AI: A New Version

By Paul Almond

This article describes a significantly revised version of the planning as modelling approach to planning in artificial intelligence (AI). This uses an AI system’s modelling system to produce probabilistic predictions of future behaviour that are equivalent to planning of future behaviour.

The previous article combined concepts from earlier articles about planning as modelling. This article now introduces a significant change to planning as modelling. The tree search approach described in earlier articles is now no longer necessary, the processing that it did being absorbed into the modelling system as part of its prediction. The model is used explicitly to make longer range predictions and the situational evaluation function score is now continually encoded as an input, the modelling system being informed of the corresponding input events, so that the modelling system can directly predict future values of the situational evaluation function score.

This new method is philosophically similar to the previous method, but provides more abstraction in planning.

Planning as modelling needs a sophisticated modelling system, but does not specify its internal workings. This article will not describe how to make a modelling system, but will show how a generic modelling system, without special features for planning, can be used to plan actions provided that it meets some criteria that would be expected in a general modelling system (being able to be informed about events as they happen and being able to make probabilistic predictions of future events).

Read full article here.