Just a quick update

It’s sunday…

So I had a bit of time. And I chose to use it to keep going a bit. I moved a couple of things, created a few more supporting functions, modified them somewhat, and added the covering part.

Covering needs to happen when for a given training instance (state, class), there is no match in the correct set.

Correct set is the subset of the Matching set for which, on top of the condition matching, the predicted class/action is also correct.

And well, it’s working:

Conclusion

Next up is “Rule Discovery”, whereby, in Michigan-style LCS, a GA-like process will take place to create offspring from two parents taken in the Correct set.

That will include cross-over (in one way or another) and mutation (with a preset pressure/probability). With these, on top of coverage, the LCS will hence create new valid rules, that will become part of the population, to be evaluated on further training instances.

Other parts of the algorithm – meant to contain the population size – will come later. :)