3 ms·

Ask HN: Human vs. Computer in Rock Paper Scizzors

3 points by EGreg 2y ago

Basically, the object is to have the human not lose by too much to a computer program in Rock Paper Scizzors (or some other random guessing game) while the computer will try to predict what the human will guess. Then have many humans together, trying to wargame against a single computer model. And then can try the same with writing text against an LLM trained on what they wrote before. The point is, that it is likely after a while humans will be extremely predictable by the model that machine learning will come up with and select competing models. Even RNNs and genetic algorithms would probably work here. Doesn’t this mean a single computer can reliably beat any group of humans in wars, trading strategies etc? Given enough battles, it will have predicted every new plan they have, eventually.

talldayo 2y ago
> The point is, that it is likely after a while humans will be extremely predictable by the model that machine learning will come up with and select competing models. Doesn't that inherently create new criteria for fitness to be judged, conditions that humans could then exploit through marginal and expected creative digression?
- leed25d 2y ago
  > ...conditions that humans could then exploit through marginal and expected creative digression? High-level professional poker players know all about this.
clipsy 2y ago
My RPS strategy is to roll a 6-sided die: if it comes up 1 or 2, I pick "Rock;" if it comes up 3 or 4, I pick "Paper;" and if it comes up 5 or 6, I pick "Scissors." Let me know when you've got an LLM that can reliably outperform me in the long run.
- EGreg 2y ago
  I should have stipulated that the humans don’t have access to an RNG. But fine, let’s say they do. In that case pick a game where completely random moves don’t win, eg poker. My point was that in a battle scenario, you can’t really surprise a computer by bluffs and feints in the long run.
  - willd13 2y ago
    When a computer plays poker, it bluffs. That is part of the optimal strategy.
  - talldayo 2y ago
    Humans always have access to RNG, whether it's psychosis or lucid innovation. We assume that people will behave the same because of regressive analysis, but we're proven wrong through digression and outlier points. Part of the "battle" is getting AI to understand and rationalize that cycle, which other humans have shown throughout history is easier said than done. You build a wall, your enemies build a ladder. You dig a moat, they come with canoes and oars. Poker indeed has a higher skill ceiling, but there is a difference between high-level and low-level play. Low-level players are cheating each other out of an understanding of the game; they win by playing smarter hands, making progressive bets, and cheating lesser players by playing smarter. Realistically you can only expect an AI to succeed at this low-level of play, since it inherently relies on a knowledge imbalance that AI could potentially exploit. But against high-level players? You're playing a different game, one that even Stockfish won't really grok. Poker masters don't behave like an AI at all, but instead like a computer. Their hand has a relative value which signals the importance of positioning a bet. That bet doesn't always have to be a bluff; sometimes it's a pressure tactic to make a smart player depend on bad cards. Multiple high-level players against each other may simply try to win by attrition, sticking to theory-optimal play to slowly grind away profitable bets. After a long enough time (or high enough value), these wins can be cashed-out apropos of nothing. How do you play smart in the long game while defensive in the short? How do you exploit bad odds and maximize favorable ones? If any of these questions were answered completely, Poker would be an obsolete game relegated to robots. But people play it, because there is indeed a game to be played once you fully understand what's on the table.
al2o3cr 2y ago
To be specific about the RPS problem, model the player and computer as two functions of [all the previous plays] that return the next play. A round is run by using each function to calculate a play based on all previous plays, then comparing the results. Now imagine you've got the "perfect" model as the computer player, that wins every time against humans. A particularly crafty human player shows up, named Cantor. His function is defined as "the choice that beats what the computer's model returns". How often does he win? This construction only works for simple games like RPS where the two players are interchangeable and a move has a well-defined "opposite".
082349872349872 2y ago
Cf https://news.ycombinator.com/item?id=36385198 https://news.ycombinator.com/item?id=36385198

3 points by EGreg 2y ago

talldayo 2y ago

> The point is, that it is likely after a while humans will be extremely predictable by the model that machine learning will come up with and select competing models. Doesn't that inherently create new criteria for fitness to be judged, conditions that humans could then exploit through marginal and expected creative digression?

leed25d 2y ago
> ...conditions that humans could then exploit through marginal and expected creative digression? High-level professional poker players know all about this.

clipsy 2y ago

My RPS strategy is to roll a 6-sided die: if it comes up 1 or 2, I pick "Rock;" if it comes up 3 or 4, I pick "Paper;" and if it comes up 5 or 6, I pick "Scissors." Let me know when you've got an LLM that can reliably outperform me in the long run.

EGreg 2y ago
I should have stipulated that the humans don’t have access to an RNG. But fine, let’s say they do. In that case pick a game where completely random moves don’t win, eg poker. My point was that in a battle scenario, you can’t really surprise a computer by bluffs and feints in the long run.
- willd13 2y ago
  When a computer plays poker, it bluffs. That is part of the optimal strategy.
- talldayo 2y ago
  Humans always have access to RNG, whether it's psychosis or lucid innovation. We assume that people will behave the same because of regressive analysis, but we're proven wrong through digression and outlier points. Part of the "battle" is getting AI to understand and rationalize that cycle, which other humans have shown throughout history is easier said than done. You build a wall, your enemies build a ladder. You dig a moat, they come with canoes and oars. Poker indeed has a higher skill ceiling, but there is a difference between high-level and low-level play. Low-level players are cheating each other out of an understanding of the game; they win by playing smarter hands, making progressive bets, and cheating lesser players by playing smarter. Realistically you can only expect an AI to succeed at this low-level of play, since it inherently relies on a knowledge imbalance that AI could potentially exploit. But against high-level players? You're playing a different game, one that even Stockfish won't really grok. Poker masters don't behave like an AI at all, but instead like a computer. Their hand has a relative value which signals the importance of positioning a bet. That bet doesn't always have to be a bluff; sometimes it's a pressure tactic to make a smart player depend on bad cards. Multiple high-level players against each other may simply try to win by attrition, sticking to theory-optimal play to slowly grind away profitable bets. After a long enough time (or high enough value), these wins can be cashed-out apropos of nothing. How do you play smart in the long game while defensive in the short? How do you exploit bad odds and maximize favorable ones? If any of these questions were answered completely, Poker would be an obsolete game relegated to robots. But people play it, because there is indeed a game to be played once you fully understand what's on the table.

al2o3cr 2y ago

To be specific about the RPS problem, model the player and computer as two functions of [all the previous plays] that return the next play. A round is run by using each function to calculate a play based on all previous plays, then comparing the results. Now imagine you've got the "perfect" model as the computer player, that wins every time against humans. A particularly crafty human player shows up, named Cantor. His function is defined as "the choice that beats what the computer's model returns". How often does he win? This construction only works for simple games like RPS where the two players are interchangeable and a move has a well-defined "opposite".

082349872349872 2y ago

Cf https://news.ycombinator.com/item?id=36385198 https://news.ycombinator.com/item?id=36385198