learn command allows training a network from training data.
As all commands in stockfish learn can be invoked either from command line (as stockfish.exe learn ..., but this is not recommended because it's not possible to specify UCI options before learn executes) or in the interactive prompt.
learn takes named parameters in the form of learn param_1_name param_1_value param_2_name param_2_value .... Unrecognized parameters form a list of paths to training data files.
It is recommended to set the EnableTranspositionTable UCI option to false to reduce the interference between qsearches which are used to provide shallow evaluation. Using TT may cause the shallow evaluation to diverge from the real evaluation of the net, hiding imperfections.
It is recommended to set the PruneAtShallowDepth UCI option to false as it will provide more accurate shallow evaluation.
It is required to set the Use NNUE UCI option to pure as otherwise the function being optimized will not always match the function being probed, in which case not much can be learned.
Currently the following options are available:
set_recommended_uci_options - this is a modifier not a parameter, no value follows it. If specified then some UCI options are set to recommended values.
targetdir - path to the direction from which training data will be read. All files in this directory are read sequentially. If not specified then only the list of files from positional arguments will be used. If specified then files from the given directory will be used after the explicitly specified files.
epochs - the number of weight update cycles (epochs) to train the network for. One such cycle is epoch_size positions. If not specified then the training will loop forever.
warmup_epochs - the number of epochs to "pretrain" the net for with warmup_lr learning rate. Default: 0.
epoch_size - The number of positions per epoch. Should be kept lowish as the current implementation loads all into memory before processing. Default is already high enough. The epoch size is not tied to validation nor net serialization, there are more specific options for that. Default: 1000000
basedir - the base directory for the paths. Default: "" (current directory)
lr - initial learning rate. Default: 1.
warmup_lr - the learning rate to use during warmup epochs. Default: 0.1.
use_draw_games_in_training - either 0 or 1. If 1 then draws will be used in training too. Default: 1.
use_draw_games_in_validation - either 0 or 1. If 1 then draws will be used in validation too. Default: 1.
skip_duplicated_positions_in_training - either 0 or 1. If 1 then a small hashtable will be used to try to eliminate duplicated position from training. Default: 0.
winning_probability_coefficient - some magic value for winning probability. If you need to read this then don't touch it. Default: 1.0 / PawnValueEg / 4.0 * std::log(10.0)
use_wdl - either 0 or 1. If 1 then the evaluations will be converted to win/draw/loss percentages prior to learning on them. (Slightly changes the gradient because eval has a different derivative than wdl). Default: 0.
lambda - value in range [0..1]. 1 means that only evaluation is used for learning, 0 means that only game result is used. Values inbetween result in interpolation between the two contributions. See lambda_limit for when this is applied. Default: 1.0.
lambda2 - value in range [0..1]. 1 means that only evaluation is used for learning, 0 means that only game result is used. Values inbetween result in interpolation between the two contributions. See lambda_limit for when this is applied. Default: 1.0.
lambda_limit - the maximum absolute score value for which lambda is used as opposed to lambda2. For positions with absolute evaluation higher than lambda_limit lambda2 will be used. Default: 32000 (so always lambda).
max_grad - the maximum allowed loss gradient for backpropagation. Effectively a form of gradient clipping. Useful for the first iterations with a randomly generated net as with higher lr backpropagation often overshoots and kills the net. The default value is fairly conservative, values as low as 0.25 could be used with lr of 1.0 without problems. Default: 1.0.
reduction_gameply - the minimum ply after which positions won't be skipped. Positions at plies below this value are skipped with a probability that lessens linearly with the ply (reaching 0 at reduction_gameply). Default: 1.
eval_limit - positions with absolute evaluation higher than this will be skipped. Default: 32000 (nothing is skipped).
save_only_once - this is a modifier not a parameter, no value follows it. If specified then there will be only one network file generated.
no_shuffle - this is a modifier not a parameter, no value follows it. If specified then data within a batch won't be shuffled.
batch_size - the number of positions per one learning step. Default: 1000
lr_step - learning rate will be multiplied by this factor every time a net is rejected (so in other words it controls LR drops). Default: 0.5 (no LR drops)
assume_quiet - this is a flag option. When specified learn will not perform qsearch to reach a quiet position.
smart_fen_skipping - this is a flag option. When specified some position that are not good candidates for teaching are skipped. This includes positions where the best move is a capture or promotion, and position where a king is in check. Default: 0.
smart_fen_skipping_for_validation - same as smart_fen_skipping but applies to validation data set. Default: 0.
max_consecutive_rejections - determines after how many subsequent rejected nets the training process will be terminated. Default: 4.
auto_lr_drop - every time this many positions are processed the learning rate is multiplied by newbob_decay. In other words this value specifies for how many positions a single learning rate stage lasts. If 0 then doesn't have any effect. Default: 0.
nn_options - if you're reading this you don't use it. It passes messages directly to the network evaluation. I don't know what it can do either.
eval_save_interval - every eval_save_interval positions the network will be saved and either accepted or rejected (in which case an LR drop follows). Default: 100000000 (100M). (generally people use values in 10M-100M range)
loss_output_interval - every loss_output_interval fitness statistics are displayed. Default: 1000000 (1M)
validation_set_file_name - path to the file with training data to be used for validation (loss computation and move accuracy)
validation_count - the number of positions to use for validation. Default: 2000.
sfen_read_size - the number of sfens to always keep in the buffer. Default: 10000000 (10M)
thread_buffer_size - the number of sfens to copy at once to each thread requesting more sfens for learning. Default: 10000
seed - seed for the PRNG. Can be either a number or a string. If it's a string then its hash will be used. If not specified then the current time will be used.
verbose - this is a modifier, not a parameter. When used there will be more detailed output during training.
bat (deprecated) - the size of a batch in multiples of 10000. This determines how many entries are read and shuffled at once during training. Default: 100 (meaning batch size of 1000000).
newbob_num_trials (deprecated) - same as max_consecutive_rejections
newbob_decay (deprecated) - same as lr_step
nn_batch_size (deprecated) - same as batch_size
use_hash_in_training (deprecated) - alias for skip_duplicated_positions_in_training
batchsize (deprecated) - same as epoch_size
use_draw_in_training (deprecated) - alias for use_draw_games_in_training
use_draw_in_validation (deprecated) - alias for use_draw_games_in_validation
convert_plain
convert_bin
interpolate_eval
check_invalid_fen
check_illegal_move
convert_bin_from_pgn-extract
pgn_eval_side_to_move
convert_no_eval_fens_as_score_zero
src_score_min_value
src_score_max_value
dest_score_min_value
dest_score_max_value
shuffle
buffer_size
shuffleq
shufflem
output_file_name