4
$\begingroup$

Today we have neural network based AI players that are comparable or better than humans in games that require extensive pattern matching and "intuition". AlphaGo is a prime example.

But these AI players usually have both neural networks and search algorithms in place. Humans, on the other hand, rely just on the pattern matching and "intuition" (even the best chess players can see just a handful of moves ahead).

So, why do AI players still require extensive search while humans don't? How would AIs like AlphaGo perform if we take the search part out?

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ > Humans, on the other hand, rely just on the pattern matching and "intuition" -- define "intuition". And how do you know that humans don't use search algorithms? It's not like we fully understand human intelligence yet.. after all, if we did, it would probably be trivial to implement "artificial intelligence" on that basis. $\endgroup$
    – mindcrime
    Feb 28 '17 at 17:29
  • $\begingroup$ @mindcrime I suppose human searches are not more than a few ply deep, at least consciously (unless it happens subsconsciously without us realizing). $\endgroup$
    – Achilles
    Mar 1 '17 at 2:13
4
$\begingroup$

It is not very accurate to say that AI players requires extensive search while humans don't. Rather, it is a question of degree.

AIs do a lot more calculating, because that's what computers are good at.

Human intuition is much more powerful than a neural network can currently hope to match, because it is much more integrated into a world of knowledge about the game, it uses orders of magnitudes more neurons and it is not a static thing that just provides a move.

But if a human player stops calculating ahead his playing strength will drop very significantly. This can be seen by assessing the performance in games with short time controls: The less time you have the less calculation is happening, your intuition however is fast.

If you take out the search component of AlphaGo it would still play quite strongly, probably at a low dan level. Of course that is also far below its strength.

So, search is always an important component of playing strength, just more so for machines.

Improve this answer
$\endgroup$
4
  • $\begingroup$ My takeaway from the few research papers I've perused is that avoiding deep search is preferred, per the extreme size of the Go, and even Chess, gamespace. $\endgroup$
    – DukeZhou ♦
    Feb 28 '17 at 20:29
  • $\begingroup$ @BlindKungFuMaster Thanks. "...it uses orders of magnitudes more neurons" - So, if AIs had a bigger neural networks, they'd be able to maintain their levels with even much shallower humanlike 2-4 ply search? $\endgroup$
    – Achilles
    Mar 1 '17 at 2:00
  • $\begingroup$ @Achilles It may help to think of it as a ratio, where, in a condition of intractability for any given processing capability, the shallower the search, the greater the number of branches that can be explored. The number of optimal plies is partly a factor of number of branches per ply, minus those branches that are pruned. I doubt there is a hard, optimal number of plies for ML Go, b/c larger data have greater utility. Thus it may be more of a question of "How many plies do I need in order to achieve the desired outcome." $\endgroup$
    – DukeZhou ♦
    Mar 1 '17 at 2:18
  • $\begingroup$ @Achilles: If you have enough data to train it, a bigger network will be better. If your network is better you will need less search for the same strength. But some details have to be calculated, not everything drops out of generalising over big datasets. And finally, 2-4 ply is a gross underestimate of how deeply decent human players calculate. In chess middlegames I usually calculate roughly 10 ply ahead. The tradeoff is more visible in search breadth. $\endgroup$
    – BlindKungFuMaster
    Mar 1 '17 at 7:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.