Nalwald

Chess engine written in Nim ♜

Nalwald is a superhuman chess engine for classical and fischer random chess. It supports the Universal Chess Interface (UCI), so it can be used with most chess GUIs, such as Arena or Cute Chess. Nalwald is written in the programming language Nim, a modern compiled systems language.

Download

git clone https://github.com/tsoj/Nalwald.git

Pre-compiled executables for Windows and Linux can be found here.

Compile

You need the Nim compiler (version 2.1.1 or higher) and the Clang compiler.

Prerequisites

nimble install [email protected]

Compiling for native CPU

nim native Nalwald

Compiling for generic CPUs

nim default Nalwald

Compiling for modern CPUs (BMI2 and POPCNT)

nim modern Nalwald

Features

• Evaluation:
  • Piece-relative piece square tables
  • 3x3 pawn structure tables
  • Piece combinations
  • Passed pawns
  • Tapered parameters
  • Optimized using gradient descent
• Search:
  • Principle variation search
  • Quiescence search
  • Transposition table
  • Move ordering:
    • Transposition table suggested best move
    • Static exchange evaluation
    • Killermoves
    • Relative and counter move history heuristic
  • Check extensions
  • Nullmove reduction
  • Late move reductions
  • Futility reductions
  • Hash result futility pruning
  • Delta pruning
  • Aspiration windows
  • Internal iterative reductions
• Multithreading support
• Supports Chess960/FRC
• Multi PV support
• UCI compatible

About

I began writing chess programs pretty much immediately after my first "Hello world!" in 2016. My first big project was jht-chess, a chess playing program with a console GUI for Linux. I used C++ but it looked more like messy C. Looking back I would say that it's hard to write worse spaghetti code than I did then, but it played good enough chess to win against amateur players. Since then, I wrote numerous chess engines, most in C++ (jht-chess, zebra-chess, jht-chess 2, squared-chess, and Googleplex Starthinker) but also one in Rust (Hactar) and now in Nim as well. While my first chess engine could barely beat me (and I am not a very good chess player, and was much less so in 2016), today Nalwald would beat Magnus Carlsen almost every time.

On this journey from an at best mediocre chess program to a chess engine that can win against the best human players, the chessprogramming.org wiki, the talkchess.com forum, and the Engine Programming Discord server have been a great source of information and motivation. At the beginning, the Wikipedia article "Schachprogramm" was really helpful, too.

Some noteworthy features of Nalwald:

• Piece-relative PSTs are piece square tables that are different depending on which square another piece is. They are added together for all pieces of all piece types.
• Futility reductions are an improvement to futility pruning. Here not only are moves skipped that are likely to be much worse than alpha. Additionally, moves that are likely slightly worse than alpha get their depth reduced accordingly to how bad they are expected to be.
• Hash result futility pruning uses hash table entries that have not a depth high enough to adjust alpha or beta, or to return a value immediately. Rather, depending on their depth, the value gets only used, when the margin to alpha or beta is big enough.
• 3x3 pawn structure tables are a way to evaluate the structure of multiple pawns. For this a 3x3 mask is used on any square for which at least two pawns (ours or enemy) fall into this mask. The pawns in that mask can be used to calculate an exact index for this structure of pawns in a 3x3 space. This index can be used to access a table, which contains values to evaluate this pawn structure. This table can then be optimized using a method like gradient descent.

Rating

Version	CCRL 40/40	CCRL 40/4	CCRL 40/2 FRC	Release Date
Nalwald
19	3306	3351	3438	July 11, 2024
18	3255	3286	3154	Aug 13, 2023
17.1	3188	–	–	June 20, 2023
17	–	3201	3051	June 5, 2023
16	2974	3001	2995	July 11, 2022
15	2899	2913	2921	Feb 8, 2022
14	2840	–	2821	Sep 16, 2021
1.12	–	–	2727	Aug 9, 2021
1.11	–	2782	–	July 22, 2021
1.10	–	–	–	July 3, 2021
1.9	2594	2627	–	June 15, 2021
1.8.1	–	2496	–	April 29, 2021
1.8	–	2416	–	April 25, 2021
Googleplex Starthinker
1.6	2357	2347	–	Aug 16, 2019
1.4	2279	2250	–	Dec 11, 2018
Squared-Chess
1.3.0	1979	2057	–	Nov 24, 2018
1.2.0	1907	1850	–	Sep 24, 2018
1.1.0	–	1800	–	Sep 20, 2018
Hactar
0.9.0	–	1351	–	Jan 13, 2018

Other commands

Generate training data

nim genData --define:almostFullCPU --run Nalwald 6_000 50_000_000 false

Create training data from PGNs

nim dataFromPGNs --run Nalwald input1.pgn input2.pgn ... inputN.pgn

Optimize evaluation parameters

nim tuneEvalParams --run Nalwald

Optimize search parameters using weather-factory

# Requires cutechess to be at /usr/games/cutechess-cli
# Values need to be updated manually in source code.
nim runWeatherFactory --run Nalwald

Run SPRT test against the master branch

# Requires cutechess to be at /usr/games/cutechess-cli
nim sprt --run Nalwald

Run bench test against commit

nim bench --run Nalwald <branch, tag or commit hash>

Run tests

nim tests --run Nalwald
nim testsDanger --run Nalwald

License

Nalwald © 2024 by Jost Triller is licensed under CC BY-NC-SA 4.0