Hacker News Reader

Comment by zimablue | Hacker News Reader

I am sceptical of Alphastar. It's not at all like alphazero, it's lots of complex parts glued together.

In the third group (second of the human and first post-review) of games, against a slightly modified version, the human won. I don't think it was a version modification, I think it was because the human had chance to review the computer's play and think of gaps. Maybe it's stronger than I think or can get stronger, but it looked like the first time a human got to sit down and think about it they could find significant holes in it's understanding. The games it won it mostly just crushed with superhuman micro, that's different in a one-off series than being consistently pro strength.

That's fair enough because it's a complex game but it just didn't seem as "polished/finished" as the chess/go demonstrations.

View on HN

I was sceptical at first but after watching MaNa story video [0] - not anymore. Starcraft2 is a very complex game, but I’ve been able to crush best blizzard AI after a week playing - yes its that weak. Seeing AlphaStar strategy, from simple correct building placement to micro decision making is so much more inspiring. And it was achieved by deep reinforcement learning - never seen a more unstable and unpredictable algorithms in my machine learning career - even in fully observable environments. What DeepMind did - achieving THAT level of play, even if AlphaStar can see the whole map and micro 3 groups of units at once - is incredible.

[0] https://www.youtube.com/watch?v=zgIFoepzhIo

You may want to try playing against some of the AIs developed by the Student Starcraft AI Tournament such as SAIDA, Locutus, CherryPi, krasi0, Iron Bot, and tscmoo.

If I could make a single recommendation to future AlphaStar opponents, it would be to prep against these AIs. AlphaStar has style overlap with the AIs mentioned above, and I'd like to see a match with human better prepared for that.

The SSCAIT has a weekly broadcast which demonstrates the strengths and weaknesses of AI players quite well.

chess/go are very different games to starcraft in terms of incomplete information, realtime nature, and micro/macro strategy. It looked very similar to the first version of OpenAI's Dota 2 bot. Their first iteration only played 1 v 1 and did so very well. The next year they expanded it to 5 v 5 and it still did very well (even though it lost to the pro teams, it beat the semi-pro caster teams).

I think skepticism is a fair reaction, but perhaps not a fair first reaction. I (personally) am optimistic about the role of AI in the future. My first reaction to the bot given its need to develop macro/micro understanding and react to incomplete information was "wow!" even if the follow up reaction was "well, its mostly god-mode micro".

Props to AlphaStar and the team that developed it.

AlphaGo started off more complex and was eventually distilled. AlphaGo started off with 2 neural networks, lots of human training games, etc, whereas AlphaZero used only 1 network and no human training games.

I imagine they're working to make it simpler and more general right now.

There was a huge version modification: the version that MaNa beat had to learn camera movement. Or do you mean that you don't think the change to require camera movement was related to the win?

I think the change was probably related -- if nothing else, time spent learning how to move the camera is time not spent learning how to e.g. build a Phoenix, which would have won that game for AlphaStar.

I don't think it was, for one thing because they mentioned that in internal testing it seemed to be equal strength and for another if you watch the game it's not some like slightly weaker strength- it's fundamentally not understanding things like how to deal with drops, shields.

it's lots of complex parts glued together

Like a human brain?

Reply on news.ycombinator.com