It is difficult to analyze the results of our training, as we did not have time to develop good tools for analyzing the produced weights as they change over time. However, we can watch games as they are played on the server and make some observations on specific things.

Early in the training session, new players will ``twitch'' a lot; that is they will continue rotating a ship in place and not do anything else. As the games advance, these players are quickly selected out. Even so, some players develop a behavior such that they twitch for a few moves, and then fire a lot.

Also, early in training players are more likely to make illegal moves, receiving error messages from the server. After about 15-20 iterations these errors disappear. We believe this is because players are penalized for this; players will quit after sending a large number of messages and the game doesn't seem to be going anywhere (we did this to prevent lock-ups during training). Players who generate a lot of errors will run out of moves faster, thereby losing games.

The A.I. does not always finish off kills. If it starts firing at a ship it won't completely sink it when it should be sure there are more segments to fire at.

We thought we'd see some weights become very large or small, depending on how useful we think they are. For instance, we expected that firing on the far half of the board would get a high weight for earlier in the game. It does not seem that particular weight is increasing much at all, although it appears to be becoming larger relative to other weights for that function for other options. Without more detailed analysis it is hard to say what is happening. However, by visual inspection it seems the A.I. always prefers to shoot on the far side of the board over shooting on any part of it.

Regardless, player behaviors, if not the weights behind them, are converging to similar patterns. Since the amplitude of mutations is rather small, this makes sense. Having a large value for one weight is only one way to produce a behavior.

Over time, games speed up. In the beginning, the A.I. players fumble around and either time out die because on player is very trigger happy. After a while, the games are over faster because the A.I. seems to be clearing the board and destroying its opponent more efficiently.

Strategies appear to fall into two categories: passive (moves a lot) and active (shoots a lot). For the most part the training seems to prefer to produce active players, but there are phases in the evolution where one can see that the passive players predominate.