Artificial Intelligence and AlphaGo's 'deep reinforcement learning'

Reading up on some of AlphaGo’s AI processes (see here) - and come across what DeepMind call deep reinforcement learning - where the computer learns by playing itself millions of times.

This strikes me as unfair. It sounds like AlphaGo is essentially building up a gargantuan database of hypothetical positions and their strengths. Humans are pretty much constrained to their previous Go games, those they’ve witnessed, and historical games they’ve researched, and recall of these is going to be very fuzzy at best. So isn’t AlphaGo gaining an unfair advantage by making up for deductive reasoning with a giant database (i.e. inductive reasoning)?

I wonder if a fairer test (closer to the essence of a Turing test) should be whether the AI is able to outplay a human based on equivalent amount of input. But I admit my understanding of brain physiology and AI are limited so this is probably conjecture. But it would be interesting to hear other views…

i don’t think ‘fair’ is a proper term in this regard. no one said that the contest between
a human and a few hundred thousand dollars worth of machinery would be fair. you wouldn’t
expect a arm wrestling contest with a front loader to be fair either.

maybe it would be better to say that this isn’t what you expect AI to be. after all, its
not a general purpose learning ‘brain’. those ‘networks’ are really very deep and wide adaptive filters, whose weights were set by running millions of trials.

i would agree with you, that its not what we think about when we say AI, at least classically. a lot of people would disagree with us.

what is really disconcerting, and amazing, is that this kind of approach was able to repeatedly dust a person who has spent their entire life studying this one topic, considered to be a particularly difficult one, and is one of the best human players to ever exist. it kind of doesn’t matter whether its fair or whether we choose to call it AI, it really changes your perspective.

I think that the million years of evolution that have shaped the human brain to be a pattern matching and learning machine are actually rather equivalent to Alphago’s input. Lee Sedol’s brain didn’t start from scratch when he was born.

1 Like

humans take a similar process. Learn the rules, then play, study other games, play for themself to memorize, play diffrent variations to figure out wich one is better and why, make assumptions and test them by playing and learn from the result.

Lee Sedol used surely more of his (Lifetime-) Resources on go then me, he played and know more games then me. Is it unfair, that he would probably ( :smiley: ) win against me?

1 Like