neuroStock is a free UCI chess engine based on the excellent Stockfish [fn:1] which is derived from Glaurung 2.1. neuroStock is an attempt to integrate a backpropagation neural network as an evaluation function.
The engine has 4 modes of play, controlled by the “Mode” UCI parameter:
0 - original Stockfish eval is logged to files. There has to be a `logs’ directory along the neuroStock executable. These log files can be processed by the neuroChessTrainer [fn:2] program to train and test networks. This mode can work correctly only on single thread so be sure to specify “Threads” UCI parameter as 1.
1 - online training. The original Stockfish eval is used to train the network specified by the “Neural Network File” UCI parameter whose evaluation is actually used for playing. This mode is the slowest but the most interesting as the engine is supposed to improve somewhat with every move (in fact, with every thought)! The neural network file format can be found at [fn:2] (not fully specified network is also allowed). If the file is missing, new network is created and will be freshly initialized. Newly trained network is saved to the same file when engine is unloaded. For this mode a single engine thread is also mandatory but there is additional “Training Threads” UCI parameter which specifies number of threads within the neural network.
2 - neural network specified by “Neural Network File” UCI parameter is used for evaluation (again not fully specified network or missing network file is also allowed but it will be freshly initialized and thus playing randomly). Multiple threads specified by the “Threads” UCI parameter are allowed.
3 - original Stockfish playing.
neuroStock is developed and thus far only tested on GNU/Linux. Instructions about building from source are identical to Stockfish’s. Invoke `make help’ within the source directory for details and options.
neuroStock is an experiment (more like meta-engine) and in combination with neuroChessTrainer[fn:2] supposed to be testing platform for different backpropagation neural networks. There are no expectations for engine strength and as far as hardly trained, it plays like a blond bimbo.