Updated chunks & rules spec

I’ve updated the chunks & rules spec to cover task synchronisation, agent messaging and a new informative section on machine learning.  See: https://w3c.github.io/cogai/chunks-and-rules.html

The test suite has likewise been extended, although it can’t be said to be full comprehensive in relation to everything in the spec. See: https://www.w3.org/Data/demos/chunks/test-suite/

In principle, machine learning for chunks and rules has two forms. The first is the automatic means by which chunk strengths are boosted when they are recalled, updated, or influenced by spreading activation, see the previous section. The second form is task-based reinforcement learning. Chunk rule engines can track which rules were used in a given thread of behavior and back propagate the reward/penalty when the task finishes successfully or unsuccessfully, respectively. If several rules match the same buffer states, then overtime this process will boost the rules that contributed to the task's success, whilst weakening rules that contributed to the task's failure.

New rules can be proposed stochastically and winnowed out through the process of trial and error. Unfortunately a purely random process for rule generation will lead to a very slow rate of learning. What we need instead is a way to use prior knowledge to guide rule generation, together with knowledge on how to decompose tasks into subtasks that the agent may already know how to accomplish, or which should be easier to learn. This is where neural networks have the potential to guide learning through their ability to handle statistical models of knowledge based upon extensive training on large corpora. A further opportunity is to use neural networks for fuzzy rules in place of explicit rules. There is lots of potential for further study!

Dave Raggett <dsr@w3.org>