Can non-conferring wisdom of xpert crowds beat AlphaGo?

[Steve Kane]

Were one to assemble a team of expert Go players, isolate them from all (especially each other) except the board and AlphaGo's moves, and have the guts to simply base the next move on majority poll, accepting that they would consciously or subconsciously "read" each other's strategy and build on it, as their moves emerged too, but still in the strict crucible of the wisdom of crowds "guess the number of sweets in the jar" methodology, but here applied to pattern manipulation and recognition, might they yet whip the beast?
Chess and Go playing practice is that where collaboration is invoved it is always based on conferring, one suspects that this can tend to dumb down. Might a "flock of genius" linked in this dangerous manner, that yet has great track record at times, make a quantum leap? Unless someone tries it, we'll never know. Go, by its very nature, lends itself to this tactic.
To level the field, the group might need to experience the game over a dozen or so games, it would be interesting to see how the nets would adapt to each other's unorthodox play. After all AlphaGo played zillions of games to limber up, would this group of agents need somehing similar? Might an injection of non-experts in related fields, ("related" might need a broad and counterintuitive definitionstructure of the game, so long they understood deeply enough the underlying structure of the game) do even better.

 

[mfb]

AlphaGo won against a team of five experts who were able to discuss. I don't see how removing the discussion would make them stronger, and larger teams with a majority decision will play very conservatively and miss innovative ideas as good ideas from individuals have no chance to be realized.
AlphaGo is far beyond human strength. So far that the developers retired it - there is no match for it apart from the program itself.

 

[Nugatory]

Were one to assemble a team of expert Go players, isolate them from all (especially each other) except the board and AlphaGo's moves, and have the guts to simply base the next move on majority poll, accepting that they would consciously or subconsciously "read" each other's strategy and build on it, as their moves emerged too, but still in the strict crucible of the wisdom of crowds "guess the number of sweets in the jar" methodology, but here applied to pattern manipulation and recognition, might they yet whip the beast?

Before taking on this question, one might want to consider a more general question: would this majority-voting ensemble play more strongly than any individual member?

 

[Steve Kane]

I think they may well start out pretty poor, because they will have to learn to work, to make decisions in a different way, perhaps being less safe and relying on the group and the system to eliminate poor individual choosing over time.

Realising that they were not playing the game alone, nor having their option overseen by "experts" should change the choosing process. What interests me is the very neurological process of choosing, and then learning. We are not looking at magic here, we are trying to tap a different kind of decision making. Failure might be as interesting as success, success would be very sweet indeed. (One could always sell the movie rights instead I suppose ;-)

I suggest one might copy the enemy, as it were, and split your very large group into two or more, and using an app perhaps, have them play rapid games against each other over some time, breed from the better groups, recombine, and after maybe a year of playing this fun game, always securing the system against conferral, mainly by rapid decision tempo, and anonymity, then you pitch your best group against an established master, depending on that you can call AlphaGo out.

The groups will be large, the tempo fast, if individuals cannot participate in every move, it will present no problem, they merely pick up where they find it again. They have to sleep anyway, and people on the other side of the planet take over. They will probably find themselves choosing in a rather different way than when playing solo, and learn to improve this, I imagine a new move every half hour or so, with a very short decision time, maybe one minute. then one stops thinking about it and gets back to work or another game.

One might suggest that at the end of each game a general discussion take place, focussing on "how the individuals felt as the group made a good choice." I don't know if this would be better within the group, or general, not knowing if you were "sleeping with the enemy".

What might be also interesting is designing apps to manage the game, or individual teams, or their breeding within the tight guidelines.

Anyway, given what already sits in the literature it seems worth a try, Ideally one could play a bunch of games against the machine with it's learning function disabled before the real match, this would level the playing field a bit.

In British military court martials, the board of three hear the evidence, then without discussion exept of points of fact and law, the most junior gives his verdict, then the second, finally the senior. The first two, if they agree, negate the senior. I doubt the five Go masters adopted such a rigourous process to avoid pitfalls and pratfalls like "Alpha deciders" "The world is watching us" and "nobody got sacked for doing what the guy we all respect would do".

For the machine it was "all in a day's work" that's when we all do our best work, trying to create this relaxation in humans is easy in a crowd.

There is little evidence that high stress helps decision making, it merely changes the part of the brain that makes the decision.

Anyway this is merely what Edward de Bono would call a "Po" - it might be ludicrous, it might spark something that leads from it, it might actually work. I would rate this at 80% - 15% - 5% with a 5% accuracy. :-) were it to work, it could raise a smile. I doubt this is prior art, I'm sure someone else is thinking about it, or has thought.

 

[Steve Kane]

I'm sure it would be pretty poor at first, designing a system to have groups play against each other via apps, breeding from groups etc will be the fun challenge.
They would have to learn to enjoy the new freedoms of choosing "out of sight" of "sharing responsibility for outcomes".

 

[fresh_42]

Before taking on this question, one might want to consider a more general question: would this majority-voting ensemble play more strongly than any individual member?

I think this is rather unlikely as in my opinion it is equivalent to the question, whether mixed strategies can result in a stronger overall strategy than an individual one. And as long as the ensemble doesn't contain players at a minor level, in which case the lack of experience can be leveled out, there won't be a benefit. At least chess games are normally lost, if conflicting strategies are pursued. The players will be busy with tactical maneuvers forced by varying strategies and a mere tactical play on one side will result which in return will lead to a lost game.

AlphaGo won against a team of five experts who were able to discuss. I don't see how removing the discussion would make them stronger, and larger teams with a majority decision will play very conservatively and miss innovative ideas as good ideas from individuals have no chance to be realized.

... is a convincing argument. How could a weaker condition lead to a better result than a stronger one? The only possibility would be a situation, in which discussions lead to a disadvantage due to social effects within the group. But this case cannot be considered here, as it meant a different set-up.

 

[Steve Kane]

From what I recall AlphaGo could barely play Go at all when it was switched on, this system, being in part based on "the enemy strategy" would be similar, but AlphaGo had no sense of fun.

 

[Steve Kane]

And chess is very different to Go https://en.wikipedia.org/wiki/Kasparov_versus_the_World
Also for the basics, although much work has yet to be done by considering participants as "agents" and using our fresh understanding of working with them to improve the management and learning process. https://en.wikipedia.org/wiki/The_Wisdom_of_Crowds

 

[Steve Kane]

Sorry, my big post above was refused by the system, and then accepted, hence my posts are a bit wonky, but I think I have brought the discussion around to the point.

 

[Steve Kane]

Personally I was unaware of the Chess example until just now, it was pure hunch, Kasparov considered it an extraordinary and exceptionally tough game with moves of great originality. Now, taken the difference between Chess and Go, and the possibility of "training the crowd", and that trying would be very cheap, it really is worth a try. Being my first post on this site, and the intense use and high regard I have for Phys.org, I am not greatly impressed with the response, from senior members too.
It is a good thing Enigma was not put before such a group, it was "proven undefeatable". Humans cracked it, chosen by a puzzle competition in part, the famous machine merely speeded things up enough to be practical.

 

[fresh_42]

Being my first post on this site, and the intense use and high regard I have for Phys.org, I am not greatly impressed with the response, from senior members too.

I'm not impressed either. E.g. nobody has claimed that chess and go aren't different. I simply talked about the difference between strategy and tactics. Chess was an example to stress the importance of these two aspects. Do you (implicitly) state, that go hasn't a strategic component? At least this would be the only plausible reason to explain your responses. I don't see that your experimental set-up is well enough described or referenced, that any complaints about the non specificity of replies would be justified.

"People will never forget how you made them feel". So true.