Starcraft AI blog

In September I looked into the source of Overkill AIIDE 2017. It has much more extensive and ambitious learning than Overkill 2016. But at the time, no replays were available, so I didn’t know how well the learning worked.

Overkill 2017 has 2 kinds of learning. One is online opening learning, which of 3 openings to play against each opponent. This is the same simple method as many other bots. The other is strategy learning, what to build/research/etc. next; I think this learning is offline, done ahead of time. The graph of win rate over time shows a modest increase in score throughout the tournament, consistent with the opening learning. The modest increase was enough to pull Overkill ahead of 5 different opponents that started out better.

Here is Overkill’s tournament history. I believe these tournaments represent 3 different versions of Overkill. The 2017 version played only in AIIDE 2017.

It’s safe to say that the learning was not successful.

What did the games look like? In every game I looked at, Overkill played sunken defense and then seemed hesitant and uncertain what to do next. It researched tech that it did not use, or did not use until much later; it made scourge when the opponent had shown no sign of air units; it added more hatcheries than its income could feed. The sunken defense worked especially poorly when the opponent went air. The basic mistakes caused weak results. I don’t know the training procedure; I can only guess that maybe the neural networks were not trained enough, or were trained on crummy data, or were misled by bugs. Or something.

Oh well. :-( Learning is hard.

Overkill-McRave is one of Overkill’s better games. #21 Overkill scored 50% against #7 McRave, which you could count as an upset. Overkill-IceBot is a more typical game with some OK action. Overkill-Juno is a sad, sad game that ran the full hour due to suicidal play by both sides.

Next: Catalog of bots which wrote learning files.