For the past few weeks, I’ve been doing extensive research on the nature of chess and the inner workings of chess AI. I think that all of the strategies that the computer uses can in one form or another be used in human play. Here is a list of some of the strategies, and a parallel to how they are used in human play:
The minimax algorithm: We can search multiple positions to a finite depth and compare the final positions and the paths taken to get there and evaluate which one maximizes our potential gain while making sure we consider the best responses our opponent has to our moves.
Monte Carlo Tree Search: When we have multiple moves that all seem to yield satisfactory results, we can use a probabilistic approach to determine which move to make. Leela Chess Zero used this approach in 2018, where she would make moves based on her probability of winning. Therefore, it would be helpful to play one type of position many times to get an idea of what moves work in the long-run and what doesn’t.
The neural network principle states that simply by playing chess, you will be able to better identify what the best moves are. Therefore, taking the time to play games and learn from your mistakes is pivotal to chess improvement.
It is these basic ideas in conjunction with consistent practice against stronger engines that will allow you to draw your games against them.
I do question how to win chess games against engines. While Magnus Carlsen, the world’s top chess player, believes he could not beat Komodo Dragon, I do believe someone can. Carlsen’s intuition — his ability to sense good moves — is very strong. However, chess engines use more mathematical techniques to calculate the best moves, and of course, this is a stronger approach than human intuition.
It’s natural to think: humans should try incorporating a mathematical approach into their chess. The most primitive approach is making a random move from the total amount of legal moves, which range from about 25 in the opening to 40 in the middlegame. This will obviously lead to sub-satisfactory results. On the other hand, AlphaZero used a probabilistic approach, using self-play to generate a vector of move probabilities that guided its search. This is a type of intuition. The difference, however, is that AlphaZero calculated tens of thousands of positions.
All in all, to beat chess engines, I would say one must improve their intuition — their fundamental understanding of chess — but to a point that is so good it even beats calculation. This would require not a “bottom-up” approach, where one gets better at calculation over time, but a “top-down” approach, where one understands chess fundamentally, in terms of its principles, and what applies when.