Ms. PacMan experiments are chugging along, but as stated in an earlier post, the score doesn’t seem to change much. But something interesting to note is the presence of general toGhost rules.
Typically, going towards a ghost is a bad idea, unless the ghost is edible, hence a toGhost rule with edible is perfect. But due to synchronous learning, the agent learns the general rule too, because the toPowerDot rule is always in effect too. This is good evidence that rules learned in parallel works, but annoying because Ms. PacMan learns less-tha-perfect rules.
What needs to happen are focused tests for finding the better of two rules.