5 ms·

On yesterday's episode of POI a character said that the Machine (an AI) has to contain a simulation of a person in order to understand what that person will do.

• visarga 10y ago
  I think AlphaGo had a pretty good simulation of its human opponent, even though strategy and intuition are hard.
  • cyphar 10y ago
    AlphaGo didn't simulate its opponent, it used the entire recorded history of Go games to extrapolate what is the best move to play in a particular situation (falling back to Monte-Carlo simulations of possible games after reducing the problem).
    • astrange 10y ago
      Game AI uses minmax decisions, so it does simulate the opponent to pick _their_ best move, then it picks the move that removes their best options. That's why it's so unsportsmanlike when losing and why it doesn't go for points when winning. It doesn't have a specific profile of the opponent, though, instead it thinks they're another AlphaGo. That's an area they could work on.
      • zardeh 10y ago
        AG doesn't actually use minimax, it uses montecarlo tree search, which doesn't even try to simulate a good opponent, it just plays lots of games very fast.
        • ylem 10y ago
          Actually, the cool thing about AG is that it doesn't just use MCTS, but that it adds in deep learning and reinforcement learning :-) I was surprised by how readable their paper was! Here's a link: http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html http://www.nature.com/nature/journal/v529/n7587/full/nature1... I think you search on the title, you can find probably find a non paywalled version.
          • zardeh 10y ago
            Yup! Their additions to the algorithm are actually super simple, it was really surprising to me, its just very computationally intensive on the front end.
        • zodiac 10y ago
          AG does use minimax in the same sense that for example stockfish does (bounded depth and with pruning)
          • zardeh 10y ago
            But that isn't minimax, its pruning by some evaluation function, where in this case the evaluation function is "blackbox Deep neural net trained on many games"
            • zodiac 10y ago
              In that case, stockfish doesn't use minimax either right, because it does pruning as well
              • zardeh 10y ago
                No, stockfish uses minimax with pruning. Minimax and MCTS are just different algorithms. MCTS simulates games randomly and creates a distribution of expected value for each one based on some cool math. It estimates the value of each state based on a sample. Minimax actually calculates them all (or all of them except for some pruned ones). That's the whole reason MCTS can work on Go and minimax doesn't. Minimax actually needs to look at all possible branches all the way down. MCTS can just say "well its been 100ms, let's use the estimates I have".
                • zodiac 10y ago
                  > No, stockfish uses minimax with pruning... Minimax actually needs to look at all possible branches all the way down. No, I assure you that stockfish definitely does not look "all the way down (ie to states where the game is over)" in most positions, even with pruning - that would take too long. It uses a static evaluation function to evaluate positions without searching down the tree further. > MCTS simulates games randomly and creates a distribution of expected value for each one based on some cool math. You described the "Monte Carlo" part of MCTS but not the "Tree Search" part. It's right in the name! See for example the first paragraph on page 20 of https://gogameguru.com/i/2016/03/deepmind-mastering-go.pdf https://gogameguru.com/i/2016/03/deepmind-mastering-go.pdf, which clearly describes AG constructing a search tree. Minimax obviously will produce perfect play in both chess and Go, but we can't use it because it takes too long. Hence we prune and truncate the search tree. When we truncate it we use an approximation of the value of the node (the static evaluation function); when we prune we sometimes use the static evaluation function as well. This is true in both MCTS and in Stockfish.
        • jheriko 10y ago
          its even more than that of course... the network has to poke its fingers in somewhere, i'd be guessing it prunes those trees - so it is a bit more than playing lots of games really fast
• gambler 10y ago
  At highest level this is true. Any prediction must be based on some model, and that model can be considered a simulation of the subject.
• SanderMak 10y ago
  That idea is beautifully explored in this episode of Black Mirror: https://en.wikipedia.org/wiki/White_Christmas_%28Black_Mirror%29 https://en.wikipedia.org/wiki/White_Christmas_%28Black_Mirro...
  • RangerScience 10y ago
    Fuuuuuuuuck that amazing show. An ounce of compassion in any of those situations and the outcome would be rather completely different... Edit: Also, goddamn dickheads wasting human upload/duplication....