Evaluation

When playing against the animat, the selection of weapons does not particularly stand out. The choices seem reasonable, although very dependent on the context. Indeed, in many cases, the animat has only a few weapons available; unconventional weapon choices are mainly caused by limited alternatives.

In addition, the fact that combat is over within a few seconds implies there is little time for errors and bugs to become obvious! To really test the weapon selection, we need to artificially prolong the fights (for instance, by making the animats invulnerable), just to check the consistency of the decisions for a longer period. Making the animats immortal also allows them to accumulate all the weapons.

From a technical point of view, we can prevent redundant computation by discarding weapons with low ammo before they are evaluated. The DT is surprisingly efficient, but every little thing helps!

A conceptual problem with this solution is that it required a lot of supervision. We needed to compute the fitness of weapons manually, or use an expert to provide examples, so that we could train the DT. Once learned, the DT has its advantages (for instance, it is general and efficient), but a more implicit approach would have been easier; we want the system to be able to learn the best weapon that improves overall performance, and not have to worry about the criteria for weapon selection.