dqn.py

from pygrn import grns, problems, evolution
import os
import argparse
import numpy as np
from datetime import datetime

parser = argparse.ArgumentParser(
    description='Evolve a GRN as a layer in a DQN model for solving tasks')
parser.add_argument('--no-learn', dest='learn', action='store_const',
                    const=False, default=True,
                    help='Turn off learning')
parser.add_argument('--no-evo', dest='evo', action='store_const',
                    const=False, default=True,
                    help='Turn off evolution')
parser.add_argument('--id', type=str, help='Run id for logging')
parser.add_argument('--seed', type=int, help='Random seed',
                    default=0)
parser.add_argument('--problem', type=str, help='Problem',
                    default='Gym')
parser.add_argument('--env', type=str, help='Gym environment',
                    default='CartPole-v0')
parser.add_argument('--steps', type=int, help='Number of training steps',
                    default=10000)
parser.add_argument('--warmup', type=int, help='Number of warmup steps',
                    default=200)
parser.add_argument('--gens', type=int, help='Number of generations',
                    default=50)
parser.add_argument('--grn_file', type=str, help='Experts from GRN file',
                    default='')
parser.add_argument('--grn_dir', type=str, help='Directory for storing GRNS',
                    default='grns')
parser.add_argument('--log_dir', type=str, help='Log directory',
                    default='logs')

args = parser.parse_args()

log_file = os.path.join(args.log_dir, 'fits_' + args.id + '.log')
pclass = eval("problems."+args.problem)
np.random.seed(args.seed)
p = pclass(log_file, args.seed, args.learn, args.env, args.steps, args.warmup)

newgrn = lambda: grns.DiffGRN()

if args.evo:
    grneat = evolution.Evolution(p, newgrn, args.id, grn_dir=args.grn_dir,
                                 log_dir=args.log_dir)
    if args.grn_file:
        with open(args.grn_file, 'r') as f:
            pcount = 0
            for g in f.readlines():
                if pcount < len(grneat.population.offspring):
                    grn = newgrn()
                    grn.from_str(g)
                    ind = evolution.Individual(grn)
                    grneat.population.offspring[pcount] = ind
                    pcount += 1
    grneat.run(args.gens)
else:
    if args.grn_file:
        with open(args.grn_file, 'r') as f:
            gen = 0
            lines = f.readlines()
            for g in lines:
                grn = newgrn()
                grn.from_str(g)
                p.generation = gen
                fit = p.eval(grn)
                with open(log_file, 'a') as f:
                    f.write('G,%s,%d,%d,%f,%d,%f,%f\n' % (
                        datetime.now().isoformat(),
                        gen, len(lines),
                        fit, grn.size(), fit, 0.0))
                gen += 1
    else:
        for i in range(20):
            grn = newgrn()
            grn.random(p.nin, p.nout, 1)
            p.generation = i
            fit = p.eval(grn)
            with open(log_file, 'a') as f:
                f.write('G,%s,%d,%d,%f,%d,%f,%f\n' % (
                    datetime.now().isoformat(),
                    i, 20, fit, grn.size(), fit, 0.0))