This month's development focus took me back to trial-and-error learning. I wanted to build on the work I'd done with the "Allergic Cliffs" puzzle game and improve Acuitas' ability to discover how the game works. I had several ideas in mind, but ended up having time to finish only one: ability to learn rules that feature negations, i.e. the absence of a feature.
 |
| Photo of a board game called Azul, by Wikimedia Commons user Gepsimos |
Previously all "rules" (action-result pairs that might represent cause-and-effect relationships) were based on commonalities between groups of actions that produced success or failure. If all attempts to put a zoombini with a green nose on the left bridge led to success, a rule such as "if a guide puts a zoombini with a green nose on a lefthand bridge, a guide succeeds" would begin to coalesce. You might also see "if a guide puts a zoombini with a green nose on a righthand bridge, a guide fails," though at easy levels of the game, the number of failures was often so low that there wasn't enough data to solidify the rule. Instead, a cluster of weaker success rules might appear. "If a guide puts a zoombini with a red nose on a righthand bridge, a guide succeeds," "If a guide puts a zoombini with a blue nose on a righthand bridge, a guide succeeds," etc. Given five color options for zoombini noses, it's pretty obvious to a human that "red OR blue OR purple OR orange" likely implies the more concise "NOT green," and I wanted Acuitas to be able to reach that conclusion as well. I also wanted complementary pairs of rules including a feature and its negation to be able to reinforce each other. (If a feature is significant, its presence and absence should matter; a pattern associated with only one of the two could be coincidental.)
So I added the ability to form rules by considering how a failure differs from prior clusters of successes. Any differences are looked at as candidates for features that should not be paired with the other features in the cluster. Further data can refine the tentative rule, excluding features that were actually irrelevant, or falsify it entirely.
To really get this to do a good job of discovering the secret rules present in a round of Allergic Cliffs, I had to stop treating individual zoombinis as "features," considering only their characteristics. This weakens the generality of the learning algorithm somewhat. In the true general case, there really could be a rule like "If a guide does not put Foozelu on a lefthand bridge, a guide succeeds" - maybe there's a quality only Foozelu has that is not part of the information available to the player. I know from experience that this game does not work that way: Allergic Cliffs rules are always based on visible properties of the zoombinis. But Acuitas isn't yet capable of the sort of meta-learning that discerns the nature of the whole game and carries over from one round to the next, so he needs a little help here. Allowing for the possibility of rules about individuals was introducing too much clutter, so I've turned it off for now.
The result is that I seem to be seeing an improvement in the robustness of rule discernment, and I often see viable rules for both bridges now, which is important for fully understanding the game and increasing one's odds of winning.
Until the next cycle,
Jenny
