Scorpio 2.8.6 MCTS neural network evaluation

[Werner]

I don´t see here if nn is used or not. perhaps I use not the correct configuration. Here my .ini file:

#############################################
# log on/off - turn logging on and off
# resign - resign value in centipawns
#############################################
log off
resign 900
##################################################################
# Hashtable sizes in mb.
# ht - main hash table shared by ALL threads
# eht - evalualtion hash table allocated PER thread
# pht - pawn hash table allocated PER thread
# Usually 1m or 2mb is enough for pawn hash table, but the others
# can be set as high as possible
##################################################################
ht 256
eht 8
pht 1
##################################################################
# montecarlo - Use montecarlo tree search (MCTS) if set to 1
# treeht - Maximum size of tree to store in memory given in MB.
# Note that this memory is not allocated at start up; it only
# specifies the maximum limit the tree could grow to.
# Note: Don't think you are making scorpio weak by not setting
# this memory to high value, infact treeht=0 is the strongest
# MCTS setting.
# frac_freeze_tree [0 - 100] - freeze tree after this percentage of time is spent
# frac_alphabeta [0 - 100] - use alpha-beta search this percent of the time
# frac_abrollouts [0 - 100] - use alpha-beta rollouts search this percent of the time
# frac_width [0 - 1e3] - determines amount of progressive widening
# mcts_strategy_depth [0 - 64] - Lookahead depth for MCTS rollouts
# alphabeta_depth [0 - 64] - Fixed search depth for standard alphabeta searches at the leaves
# evaluate_depth [0 - 64] - Fixed search depth for evaluation of new children during expansion
# backup_type [0 - 5] - minmax = 0, averaging = 1, mixed = 2, for mixing score with previous ID use 3,4,5 resp.
# reuse_tree - reuse MCTS tree between moves
##################################################################
montecarlo 1
treeht 12800
frac_freeze_tree 100
frac_alphabeta 100
frac_abrollouts 20
frac_width 100
mcts_strategy_depth 30
alphabeta_depth 16
evaluate_depth 0
backup_type 0
reuse_tree 1
#############################################################################
# Parallel search setup
# mt - number of processors
# smp_depth - split depth for smp
# cluster_depth - split depth for cluster search
# message_poll_nodes - number nodes to search before checking for message
#############################################################################
mt 1
smp_depth 8
cluster_depth 12
message_poll_nodes 20
smp_type YBW
cluster_type YBW
##########################################################################
# EGBB
# egbb_path - path to endgame egbbs
# egbb_cache_size - cache size for egbbs
# egbb_ply_limit_percent - percent of plies from the root to egbb probing
# egbb_depth_limit - depth from the leaves to egbb probing
# egbb_load_type - egbb load type with the following values
# 0 = none are loaded
# 1 = all 3/4 men are loaded
# 2 = Not implemented yet
# 3 = all 5 men are loaded
##########################################################################
egbb_path c:\negbb64\
egbb_cache_size 32
egbb_load_type 3
egbb_depth_limit 6
#######################################
# Path to neural network
#######################################
use_nn 1
nn_path c:\Users\xx\Arena\Engines\ScorpioNet\nets\net-6x64.pb


use_nn is Default set to 0 ??

[CMCanavessi]

If I set use_nn 1 in the .ini file, scorpio automatically stops responding when it starts "thinking". If I set it to 0 it works just fine, but I don't know if it uses the network file or not.

[Milos]

I don't think you understand how MCTS works. In Scorpio it is using just value network as direct rollout replacement (i.e. as eval). Policy network is not used to guide the search. MCTS is completely agnostic to the way evaluation only is performed (A/B rollout, handcrafted NN eval, LC0/A0 NN eval, etc.). Format, number of layers, input planes of NN and all the internals are irrelevant for search.

Except....the value network is an output of the neural network, and is simply a series of transformations of the inputs and the weights. So of course the format of the neural network matters! If you don't have the same format, then the value network that gets output is junk. How MCTS uses that to guide the search is what is irrelevant, since it's garbage information.

That is ofc true, but implementing the same format as in LC0 is rather trivial. Daniel used TF and he could have just copied TF interface from LC0 from Alex for LC0 NN. But you are right, I checked to source code and he didn't, so Scorpio doesn't support LC0 NN.

[Milos]

If I set use_nn 1 in the .ini file, scorpio automatically stops responding when it starts "thinking". If I set it to 0 it works just fine, but I don't know if it uses the network file or not.

If you wanna use NN you set use_nn to 1, but you should also have full TF installed/compiled for your target OS and in path for Scorpio to find it.

[Werner]

you should also have full TF installed/compiled for your target OS and in path for Scorpio to find it

… seems to be quite complicated for me (for Windows 10 x64)

[Milos]

you should also have full TF installed/compiled for your target OS and in path for Scorpio to find it

… seems to be quite complicated for me (for Windows 10 x64)

Yes building TF for C/C++ is quite a pain on Windows. Seems Google doesn't quite like or care providing support for MS OS's beside python.
If you still care to do it yourself, here is probably the best tutorial:
https://medium.com/@shiweili/building-t ... 9c90e23e6e

[Albert Silver]

you should also have full TF installed/compiled for your target OS and in path for Scorpio to find it

… seems to be quite complicated for me (for Windows 10 x64)

Yes building TF for C/C++ is quite a pain on Windows. Seems Google doesn't quite like or care providing support for MS OS's beside python.
If you still care to do it yourself, here is probably the best tutorial:
https://medium.com/@shiweili/building-t ... 9c90e23e6e

Do you mean that building an NN with TF is a pain in Windows? Or that providing support for a TF-built NN is hard?