We do the same with our chess games as a sequence of moves. In this task, the model “learns” to find the best possible vector representations for the words. A “window” of a certain length around a word is considered. To find these vectors, neural networks are trained to predict either a word from the surrounding words or the surrounding words from the word. In the pre-processing, I have chosen to represent the games in such a way that a game corresponds to a sentence and each move corresponds to a word.

In our example, moves in chess are considered similar if they often occur before or after the same moves in games. Vectors are chosen in such a way that words that often occur in the same context in texts are considered similar and are therefore assigned vectors that have a small distance between them. In order to be able to calculate a similarity between moves, we use an idea from Natural Language Processing in which words are represented by vectors. In this example, the top 150 moves were selected. Pgn-extract.exe lichess_db_standard_rated_2016–06.pgn -o2016–06.pgn -bl10 -w1000 - noresults - notags - nomovenumbers - nocomments - quietĪll games from July 2016 are pre-processed. With the command line program “ pgnextract” all comments and move numbers were removed and each game was written into one line of a text file. To be able to use the data for my purposes, I did some pre-processing. If you choose Moves (PGN), the moves and game information are saved as plain. The moves in the files are given in Standard Algebraic Notation. If you choose Chess Game, you can open the file in Chess to continue the game. There are millions of games per month, which allows me to do my analyses. I recommend just picking one parameter to follow. Move A might have the most human games, move B might be the engine's choice, move C might have the best score, move D might be recommended in a book, move E might be the choice of top players. These are simple text files with a defined structure, which on the one hand can be read and written by almost any chess software, and on the other hand, can be easily manipulated as text files. Picking the right opening move is very difficult. At you can find all the games played on this server collected by month for download in Portable Game Notation (PGN) format. The source for my analyse s are files of games played on the internet chess server Lichess. Is it possible to recognize a general structure and to represent it visually? Data source If two moves often occur after or before the same moves, then these moves are similar in a certain sense.įor example, which move is close to the opening with the queenside pawn “d4”? In this article, I want to analyze which moves in a game of chess are close, in the sense that they often occur in similar situations in games. Similarity map of chess moves / Image by Author

Pgn chess opening moves