Google's AI defeats a top go player

7 posts (showing 1-7)
Kryog

Market Level 0Community Level 1
0 posts

There are probably many people in here who already know about the ancient chinese game of go, it's a classic game among the 'nerdy' type and has been for a few millennia. For those who don't: it's an abstract strategy board game, where you surround territories, and it's a lot more complicated than it sounds, deeper than chess, has many professional players in Asia, etc. really you should look it up it's pretty damn fun.

These days, human and machine are fighting for supremacy in go that used to be impossible for computers to play properly at even a decent amateur level, because raw computer power cannot evaluate the absurd breadth of possible combinations in this game to any significant depth, something that a human player does by reasoning in terms of 'intuition' and 'shape' for the most parts.

It turns out that one if not the best professional go player of our generation has now been repeatedly and spectacularly destroyed by an AI called AlphaGo and developed by DeepMind (a company bought by Google), using a new approach based on neural networks.

I like to imagine that the next step is making AI beat human players in games like Starcraft II (currently even on the hardest non-cheating difficulty, and despite their obvious APM advantage, bots are really no match for a veteran human player).

Also you could, in a not so distant future, harness the power of AI not just to crush humans, but to create video game experiences (say, levels in a platformer) specifically optimized to fulfill the observed needs of each player, based on realtime feedback, in a way that is more sophisticated and pleasing than any game design approach based on the current fixed heuristics of human game design could achieve. Think Super Mario Maker, with an AI creator that spits out every time artfully convoluted levels for everyone to enjoy. Not just a random procedural generation whose fixed parameters are injected by the programmer once and for all, but the computer having via a neural network a 'real' adaptable understanding of the underlying concepts of game design, so he can create fresh content indefinitely, or maybe entire games.

So, any thoughts on Lee Sedol being beaten by AI in such a sudden and unexpectedly dominant way? Is there any game or intellectual domain or knowledge-based activity that is safe from AI, now that chess, jeopardy, and go are out of reach, in ten or twenty years will we still be making games ourselves or let the AI generate them in superior ways with only assets and a game genre as inputs ?

posted 2016-03-13T22:03:04-07:00 | edited 2016-03-13T22:06:55-07:00
Blackizzy

Market Level 1Community Level 5
421 posts

 Well i don't know too much about go, but i used to play chess intensively for nearly half a decade, and stopped as  i was pretty deceived to see people use bot to play against human, only to pretend they are smarter. It reduced this magnificent game to a pile of calculations. I tried to play against shredder and the best i can do is draw whit the soft being run on moderate calculation. Forget about the last level, the best Sicilian defense which proved to be effective for centuries was destroyed within 10 moves.

  For the starcraft players, i presume it's really a matter of time before machine could kick pro asses. They don't have to care about their focus level and the time it takes to select units, use spells and give instructions is already winning over human. I bet if a company had to offer 100,000$ USD to the best starcraft AI, it would whip the pros within 6 months from now. I am not even a good coder nor mathematician, but i would give it a try, as it sounds pretty interesting as a challenge.

   In any case, for me, the real fun is to play against human, friend or not.  Few are sad because the digital calculator piles number faster than a human. I feel the same about a computer, seeing it like an ally rather than an opponent.

posted 2016-03-14T15:09:11-07:00
FoumartGames

Market Level 3Community Level 4
209 posts

Thinking of AI our goal as game creators is to make the AI capable of beating even the smartest / most skilled of human. That would allow us to make the AI able to reduce it's difficulty in order to provide the ultimate level of fun for all kind of players.

The impression after a played game must be as if the CPU was a real opponent, making bluffs and sacrifices, making the player believe he's playing against a real human. lol

There is a site HackerRank that provides all kind of challenges in programming. Artificial Intelligence challenges for example are to build a bot that plays a certain game (like Battleships) and make it beat others bots - it's really fun - try it: https://www.hackerrank.com/challenges/battleship
If interested I can post you a startup code in JS.

posted 2016-03-15T09:33:26-07:00
Blackizzy

Market Level 1Community Level 5
421 posts

 Thanks for this info FoumartGames, it's very nice challenges featured there. I may try one of those someday, just need to find time and courage !

posted 2016-03-15T19:06:00-07:00
ozdy

Market Level 6Community Level 13
1482 posts

I find it hard to believe a game like Go was hard to make an algorithm for previously that couldn't beat a top human. Computers are much more capable in calculations than the human brain, and the state of a Go position, while big, is very easily representable for a computer program.

posted 2016-03-16T13:44:38-07:00
Andromedus

Market Level 7Community Level 5
564 posts

Well, either all the people who tried to program AI for Go before now have been really bad programmers, or it's actually quite hard!

posted 2016-03-16T16:23:21-07:00
ozdy

Market Level 6Community Level 13
1482 posts

Btw, anybody with some hands on experience with neural network knows what basic stuff they are and they only depend on a gazillion of input-output pairs before being able to process a new input.

posted 2016-03-16T16:37:22-07:00 | edited 2016-03-16T16:37:49-07:00