4 ms·

A bit offtopic: is it still possible to create a Chess/Go type game that is easy for humans but hard for AI? Or are we past that point already?

by krzat 7y ago

shagie 7y ago
The game Arimaa attempted to do that. https://en.m.wikipedia.org/wiki/Arimaa https://en.m.wikipedia.org/wiki/Arimaa
- V-2 7y ago
  Is it more computationally complex than Go?
  - shagie 7y ago
    This game is played with chess pieces on a chess board... though they have different meanings. Each player sets up their pieces in any way they want on the back two rows. The number of possible of possible starting positions for each player is 64,864,800. This means that the starting game position possibility is 4.2e15. Next is the branching factor. Chess is about 35 possible moves each play for each player. Arimaa, each player makes two successive moves. This makes for a very large branching factor... 17,281 possible moves each turn ( http://arimaa.janzert.com/bf_study/ http://arimaa.janzert.com/bf_study/ ). Go has a branching factor of 250. https://en.wikipedia.org/wiki/Game_complexity https://en.wikipedia.org/wiki/Game_complexity has a chart on it... Go has a larger board size and space state complexity, but Arimaa has a larger game tree size and branching factor. That 35 and 17,281 thing... 35^8 ~= 17,000^3. If you have a computer program that can search all the possible moves 8 deep in chess (Deep Blue went 6-8 typically), it could only look 3 deep into Arimaa (a similar search space is 5 deep in Go).
halfarmbandit 7y ago
Probably no. The issue with tabletop games is that we can model them with ease. This allows bots to play myriads of games and the ability to look ahead in possible moves.
plopz 7y ago
I could see a game like Mao being difficult for AI or really any game with an inconsistent ruleset. https://en.wikipedia.org/wiki/Mao_(card_game) https://en.wikipedia.org/wiki/Mao_(card_game)
- UncleMeat 7y ago
  There is actually an entire field of general game playing that can be extended to receiving partial rulesets at the start of the game.
bo1024 7y ago
It's not clear what it means for a two-player game to be "easy for humans but hard for AI" -- but if you mean that humans can pretty easily beat AI, probably the way to go is a game of incomplete information. For example in hold'em poker, AI have only been able to beat humans in the last 1-2 years.
humble_engineer 7y ago
They can beat us at any type of game that doesn't require split second thinking.
yifanl 7y ago
Sure, a game where the rules can't be easily codified but understandable by humans would pose a good challenge. I'd like to see a computer play Calvinball :)
alach11 7y ago
Humans still reign supreme (for now) at games which emphasize a heavy strategic component with a very large search space. For example, the OpenAI Five can only beat humans at DOTA with a very restricted scope (restricted champions and items).

A bit offtopic: is it still possible to create a Chess/Go type game that is easy for humans but hard for AI? Or are we past that point already?

by krzat 7y ago

A bit offtopic: is it still possible to create a Chess/Go type game that is easy for humans but hard for AI? Or are we past that point already?

shagie 7y ago

The game Arimaa attempted to do that. https://en.m.wikipedia.org/wiki/Arimaa https://en.m.wikipedia.org/wiki/Arimaa

V-2 7y ago
Is it more computationally complex than Go?
- shagie 7y ago
  This game is played with chess pieces on a chess board... though they have different meanings. Each player sets up their pieces in any way they want on the back two rows. The number of possible of possible starting positions for each player is 64,864,800. This means that the starting game position possibility is 4.2e15. Next is the branching factor. Chess is about 35 possible moves each play for each player. Arimaa, each player makes two successive moves. This makes for a very large branching factor... 17,281 possible moves each turn ( http://arimaa.janzert.com/bf_study/ http://arimaa.janzert.com/bf_study/ ). Go has a branching factor of 250. https://en.wikipedia.org/wiki/Game_complexity https://en.wikipedia.org/wiki/Game_complexity has a chart on it... Go has a larger board size and space state complexity, but Arimaa has a larger game tree size and branching factor. That 35 and 17,281 thing... 35^8 ~= 17,000^3. If you have a computer program that can search all the possible moves 8 deep in chess (Deep Blue went 6-8 typically), it could only look 3 deep into Arimaa (a similar search space is 5 deep in Go).

halfarmbandit 7y ago

Probably no. The issue with tabletop games is that we can model them with ease. This allows bots to play myriads of games and the ability to look ahead in possible moves.

plopz 7y ago

I could see a game like Mao being difficult for AI or really any game with an inconsistent ruleset. https://en.wikipedia.org/wiki/Mao_(card_game) https://en.wikipedia.org/wiki/Mao_(card_game)

UncleMeat 7y ago
There is actually an entire field of general game playing that can be extended to receiving partial rulesets at the start of the game.

bo1024 7y ago

It's not clear what it means for a two-player game to be "easy for humans but hard for AI" -- but if you mean that humans can pretty easily beat AI, probably the way to go is a game of incomplete information. For example in hold'em poker, AI have only been able to beat humans in the last 1-2 years.

humble_engineer 7y ago

They can beat us at any type of game that doesn't require split second thinking.

yifanl 7y ago

Sure, a game where the rules can't be easily codified but understandable by humans would pose a good challenge. I'd like to see a computer play Calvinball :)

alach11 7y ago

Humans still reign supreme (for now) at games which emphasize a heavy strategic component with a very large search space. For example, the OpenAI Five can only beat humans at DOTA with a very restricted scope (restricted champions and items).