How does a game of chess work between two top level players? Do they have to think like 25 moves ahead to find any sort of opening against the opponent?
In: 347
Disclaimer: I’m really not very good at the whole “explaining like you’re *actually* five” thing, but I do have a lot of subject knowledge and expect that someone who is at least a high schooler should be able to understand my explanation without any issues. Let me know if you have questions though.
Let’s start with a simpler game, tic tac toe. Most of us learn pretty quickly that as long as X goes in the corner or middle first, it’s very simple to force a draw, and even if you don’t go through all the options exhaustively, you probably go through a lot of them just by playing enough times. This being said, what you’re doing when you think to yourself “if I make move A, they have moves B and C; if they make move B, I have move D which lets me set up a 2 way trap to win, and if they make move C I can block them and force a draw, so move A leads to a draw” is something called the *minimax algorithm*, which is just a fancy way of saying that you evaluate the effectiveness of a move based on the best result that you can guarantee (move A led to a draw because your opponent can choose between a draw or a win for you, whereas a move in which you have a path to victory for any move they make would be a winning move, and a move in which they have a path to victory you cannot prevent is a losing move).
Chess can technically be completely solved in this manner—if you were to look at every possible terminal position (a position where the game is over), and worked backwards to evaluate every move leading up to these terminal positions, you would successfully solve chess completely and have everything you need to play it optimally. This would take orders of magnitude more computing power than is currently possible, however, and so instead we have to do our best to imitate the moves that this kind of analysis would lead to, without having the information necessary to verify that your move forces a win (or at least prevents a draw).
Luckily, humans have gotten pretty good at evaluating positions qualitatively. We understand that generally having more pieces than the opponent is an advantage, as well as having a king in a safer position, having connected pawn chains, having passed pawns close to promoting, having bishops over knights on a relatively open board, etc and from experience can get pretty good at using this understanding to make moves that we think will allow us to maintain or gain an advantage later on, even if we can’t necessarily give the exact moves well into the future that will allow for this. We can also understand the concept of *tactics*, relatively short sequences of moves that allow us to force a position we see as favorable, such as a sequence of moves that wins a rook without making any visible concessions, and generally tactics follow directly from a good positional understanding of the game—if your pieces are well-positioned, the threats they can make are easily accessible and restrictive on your opponent’s moves.
Computers take this a step further using a variety of methods, with most modern chess engines using neural networks. Remember the minimax algorithm I described? Well, one approach we can take (which is the approach used by Stockfish, the strongest engine) is that if a computer can evaluate a position by assigning it a number, where that number represents whom the computer thinks will win with what level of certainty (say, +100 is a guaranteed won game, -100 is a guaranteed lost game, 0 is a draw, and something like +10 is a position we think is good, but can’t compute far enough forward to guarantee that it’s good), then we can use those number assignments to find what our best move is, based on what sequence of moves guarantees the highest evaluation possible to us. Note that we’re not hoping our opponent stumbles into a position that’s good for us, we are, to the best of our ability, looking to guarantee a position that’s good for us, and doing so well is dependent on 1) being good at assigning that numerical evaluation to a position, and 2) being able to look as far into the future as possible.
Latest Answers