Starcraft AI blog

A few unrelated things I’ve noticed on SCHNAIL.

bad bots

There are a couple of bad bots that don’t start up. One is named “test”. Since the starting elo of 1500 is near the top elo for bots, and the elo falls slowly since nobody plays many games against a bot that doesn’t work, in ranked play the bad bots tend to be matched against the stronger humans. That’s unfortunate.

human habits

In practice games, most humans seem to have favorite bot opponents, and don’t experiment widely. That’s interesting, and to me a little surprising. Most pick opponents that they usually lose to. I think either they are truly practicing, or else they’re taking it as a challenge. The strongest humans don’t have bots that they usually lose to, though.

I have the impression that there have been more ranked games over time, as a proportion of all games. Maybe players are getting familiar enough to feel comfortable with it. Many humans choose practice games every time, though, which is of course a perfectly good choice.

It’s curious, but human games are more stereotyped and less varied than bot games. At least at this level on SCHNAIL. (It’s not so at the highest level of human play.) Most humans on SCHNAIL seem to play the same base strategy every game, like the bot SAIDA does, with adaptations to what they’ve scouted.

Humans quit the game before any fighting in a surprising (and irritating) proportion of games. It’s impossible to know how much it’s due to “Oops, I messed up my split, try again,” versus “I’d better take this call,” versus “ack, why’d it freeze now?” From my point of view as a bot author, these are noise games that make the data harder to interpret.

Steamhammer’s play

When Steamhammer wins, it’s not due to better strategy (it’s worse than humans), or better tactics (it’s far, far worse), or better micro (better in some aspects, but overall worse). It’s occasionally due to a rush, zerglings or lurkers breaking in by surprise, but not often; humans are better at adapting and more resilient in defending than bots. Usually it’s due to better macro: “I lost this base, I lost that base, I didn’t kill any bases of yours, but I’m still ahead in workers. Say goodnight.”

Steamhammer’s inability to foresee an enemy army’s intent is a severe weakness. “Your army is over there, near my expansion? It’s out of position, I’ll move right up in front of your natural. Wait, you’re attacking my expansion now? Run run run to defend! Oops, too late.”

In games versus terrans, it’s clear that Steamhammer is now stronger with lurkers than with mutas. Lurker play is still crude, but the recent improvements have paid off.

gg when losing

As mentioned in the post the HumanOpponent flag, Steamhammer behaves a little differently when facing a human opponent. In particular, it gives up much earlier when losing. The rule is explained in the old post.

When I wrote that surrender rule, I doubted that it was tuned well. I have been pleasantly surprised, because it works great. In actual games, Steamhammer only surrenders when it is truly losing, and the timing is almost always reasonable: Late enough that it is also clear to the opponent that Steamhammer is losing, and early enough not to waste too much human time. I wish everything worked that well!

Next: I played around a bit with the SCHNAIL api. It’s easy to extract data in JSON format. I want to analyze it a bit.