(May 27th, 2023, 14:36)GMBarak Wrote: IMO if the same development time that was put into alpha zero would be put into CoM AI, no human would be able to win at fair difficulty. unless the human had extraordinary luck.
Problem for me with Master is that if I get lucky I can snowball it into a trivial win.
if you don't get lucky in Master even if you are good player and get attacked by Chaos wizards' herd of Gargoyles or Nature wizard herd of giant spiders... you lost no matter how good you are.
Obviously the AI would play far better than now if they would have put in that much time, money, and effort as they put into developing chess AI and Go AI.
But I disagree that it would play nearly at the human expert level due to the complexities and the uncertainties involved in playing COM compared to the very simple and deterministic rules, the very low number of options to choose from for every move (average branching factor), the very short games of both Chess and Go, the full access to the game state at all times, and for those games being only 2 person games leaving out diplomacy.
Brute force, heuristics, weight functions, and neural networks that they used for Chess and Go AIs to make them world Champions will not work that well for COM. Knowledge representation and processing element (neuron) association are far more complex in COM and leaving the Machine Learning AI to figure it out alone without prior knowledge that works so well in AlphaGo Zero is not likely to produce results that fast or that strong.
On this I disagree with Elon Musk too that their existing OpenAI would quickly learn any game to the expert level. If it can do that, why did they not demonstrate it yet for a complex game like this?
I also disagree that winning on Master is not possible with bad luck. There are various strategies to compensate for bad luck, but of course there are situations the luck is so bad that it really cannot be won vs the AI.