batch of 100 games at 1+0.01 tc
I will converge ... but take your time ...
Code: Select all
#!/usr/bin/python
# -*- coding: utf-8 -*-
from subprocess import Popen, PIPE
import sys
import platform
import random
import math
import multiprocessing
from joblib import Parallel, delayed
import pyswarms as ps
from random import randrange
from pyswarms.utils.plotters import plot_cost_history
from random import randrange
import matplotlib.pyplot as plt
import numpy as np
from termcolor import colored
threads = 8
ngames = 100
if platform.system() == 'Windows':
cutechess_cli_path = "\"C:\Program Files (x86)\cutechess\cutechess-cli.exe\""
else:
cutechess_cli_path = "/ssd/cutechess/projects/cli/cutechess-cli"
engine = 'conf=minic_dev_uci'
engine_param_cmd = ' option.{name}={value}'
opponents = [ 'conf=minic_dev_uci' ]
options = '-each tc=1+0.01 -openings file=/ssd/Minic/Book_and_Test/OpeningBook/Hert500.pgn format=pgn order=random plies=24 -draw movenumber=80 movecount=5 score=5 -resign movecount=5 score=600 -pgnout out.pgn'
params = [
["PawnValueMG" , 0 , 2000, 500 ],
["PawnValueEG" , 0 , 2000, 500 ],
["KnightValueMG", 0 , 2000, 500 ],
["KnightValueEG", 0 , 2000, 500 ],
["BishopValueMG", 0 , 2000, 500 ],
["BishopValueEG", 0 , 2000, 500 ],
["RookValueMG" , 0 , 2000, 500 ],
["RookValueEG" , 0 , 2000, 500 ],
["QueenValueMG" , 0 , 2000, 500 ],
["QueenValueEG" , 0 , 2000, 500 ],
]
def run_one(i,fcp,scp,ngames):
if i % 2 != 0:
fcp, scp = scp, fcp
cutechess_args = '-engine %s -engine %s %s' % (fcp, scp, options)
command = '%s %s' % (cutechess_cli_path, cutechess_args)
print("Running game {}/{} \r".format(i+1,ngames), end='')
process = Popen(command, shell = True, stdout = PIPE)
output = process.communicate()[0]
if process.returncode != 0:
sys.stderr.write('failed to execute command: %s\n' % command)
return None
result = -1
for line in output.decode("utf-8").splitlines():
if line.startswith('Finished game'):
if line.find(": 1-0") != -1:
result = i % 2
elif line.find(": 0-1") != -1:
result = (i % 2) ^ 1
elif line.find(": 1/2-1/2") != -1:
result = 2
else:
sys.stderr.write('the game did not terminate properly\n')
return None
break
if result == 0:
return 1
elif result == 1:
return -1
elif result == 2:
return 0
def run(ngames,threads,fcp,scp):
results = Parallel(n_jobs=threads)(delayed(run_one)(i,fcp,scp,ngames) for i in range(ngames))
l = list(results)
draws = l.count(0)
wins = l.count(1)
losses = l.count(-1)
mu = max(0.00001,wins + draws/2)/ngames # avoid div by zero
elo = -math.log(1.0/mu-1.0)*400.0/math.log(10.0);
los = .5 + .5 * math.erf((wins-losses)/math.sqrt(2.0*(wins+losses)));
print("\nwins {}".format(wins))
print("draws {}".format(draws))
print("losses {}".format(losses))
print(colored("elo {}".format(elo), 'red' if elo<0 else 'green' if elo > 15 else 'yellow'))
print("LOS {}".format(los))
return sum(results),los,elo
def obj_func_one(x):
print("\nCurrent params {}".format([int(xx) for xx in x]))
# configure engine
fcp = engine
scp = opponents[random.getrandbits(16) % len(opponents)]
for i in range(len(params)):
option = engine_param_cmd.format(name = params[i][0], value = int(x[i]))
fcp += option
return ngames - run(ngames,threads,fcp,scp)[0]
def obj_func_many(x):
res = [None]*len(x)
for k in range(len(x)):
res[k] = obj_func_one(x[k])
return res
def PSO():
# Set-up hyperparameters
options = {'c1': 0.5, 'c2': 0.3, 'w':0.9}
# Call instance of PSO
optimizer = ps.single.GlobalBestPSO(n_particles=15, dimensions=len(params), options=options, bounds=([params[i][1] for i in range(len(params))], [params[i][2] for i in range(len(params))]))
# Perform optimization
best_cost, best_pos = optimizer.optimize(obj_func_many, iters=1000)
plot_cost_history(optimizer.cost_history)
plt.show()
def main(argv = None):
if argv is None:
argv = sys.argv[1:]
PSO()
if __name__ == "__main__":
sys.exit(main())
Code: Select all
#!/usr/bin/python
# -*- coding: utf-8 -*-
from subprocess import Popen, PIPE
import sys
import platform
import random
import math
import pyswarms as ps
from pyswarms.utils.plotters import plot_cost_history
from random import randrange
import matplotlib.pyplot as plt
# parameters to optimize
params = [
["staticNullMoveMaxDepth0", 0, 15, 6],
["staticNullMoveMaxDepth1", 0, 15, 6],
["staticNullMoveDepthCoeff0", 0, 300, 80],
["staticNullMoveDepthCoeff1", 0, 300, 80],
["staticNullMoveDepthInit0", 0, 300, 0],
["staticNullMoveDepthInit1", 0, 300, 0],
]
def obj_func_one_fake(x):
return sum([ (x[i]-params[i][2]/2)*(x[i]-params[i][2]/2) + randrange(-100,1000)*params[i][2]/100 for i in range(len(x))])
def obj_func_many(x):
res = [None]*len(x)
for k in range(len(x)):
res[k] = obj_func_one_fake(x[k])
return res
def PSO():
# Set-up hyperparameters
options = {'c1': 0.5, 'c2': 0.3, 'w':0.9}
# Call instance of PSO
optimizer = ps.single.GlobalBestPSO(n_particles=15, dimensions=len(params), options=options, bounds=([params[i][1] for i in range(len(params))], [params[i][2] for i in range(len(params))]))
# Perform optimization
best_cost, best_pos = optimizer.optimize(obj_func_many, iters=1500)
plot_cost_history(optimizer.cost_history)
plt.show()
def main(argv = None):
if argv is None:
argv = sys.argv[1:]
PSO()
if __name__ == "__main__":
sys.exit(main())
