def run(arguments):
parser = argparse.ArgumentParser(description='Perform linear GP evolution for a given environment.')
# Program configuration
parser.add_argument('--minps', dest='min_prog_size', type=int, help='Minimum number of instructions per Program', default=32)
parser.add_argument('--maxps', dest='max_prog_size', type=int, help='Maximum number of instructions per Program', default=1024)
parser.add_argument('--padd', dest='padd', type=float, help='Instruction addition strength', default=0.7)
parser.add_argument('--pdel', dest='pdel', type=float, help='Instruction deletion strength', default=0.7)
parser.add_argument('--pmut', dest='pmut', type=float, help='Instruction mutation strength', default=0.7)
# Trainer configuration
parser.add_argument('--generations', dest='num_generations', type=int, help='Number of generations over which evolution is performed', default=50)
parser.add_argument('--pop', dest='population_size', type=int, help='Learner population size', default=200)
parser.add_argument('--keep', dest='percent_keep', type=float, help='Percentage of surviving Learners', default=0.3)
parser.add_argument('--fast', dest='fast_mode', type=bool, help='Skip some re-evaluations', default=True)
parser.add_argument('--skips', dest='num_skips', type=int, help='Number of generations over which to skip re-evaluation', default=3)
parser.add_argument('--episodes', dest='num_eps_per_gen', type=int, help='Number of episodes over which an agent is evaluated each generation', default=3)
parser.add_argument('--verbose', dest='verbose', type=bool, help='Do print out info to the console during evolution', default=True)
parser.add_argument('--agent', dest='agent_save_name', type=str, help='Name under which to save the evolved agent', default="")
# Environment configuration
parser.add_argument('--env', dest='env', type=str, help='OpenAI environment', default="CartPole-v1")
parser.add_argument('--statespace', dest='statespace', type=int, help='Length of flattened state space', default=4)
args = parser.parse_args(arguments)
if args.env != "CartPole-v0" and args.env != "CartPole-v1":
print("Woops! So far this module only works in the CartPole environment!")
return
if args.statespace != 4:
print("Woops! So far this module only works in the CartPole environment with a statespace size of 4!")
return
ConfigureProgram(
num_inputs = args.statespace,
min_prog_size = args.min_prog_size,
max_prog_size = args.max_prog_size,
p_add = args.padd,
p_del = args.pdel,
p_mut = args.pmut)
ConfigureTrainer(
num_generations = args.num_generations,
population_size = args.population_size,
percent_keep = args.percent_keep,
fast_mode = args.fast_mode,
max_num_skips = args.num_skips,
num_eps_per_gen = args.num_eps_per_gen,
verbose = args.verbose,
agent_save_name = args.agent_save_name)
env = gym.make(args.env)
trainer = Trainer(env)
trainer.evolve()
def run(arguments):
if len(arguments) == 0:
print("ERROR - No arguments given to main")
sys.exit(0)
# Setup the command line parsing to read the environment title
parser = argparse.ArgumentParser(
description='Perform TPG evolution for a given environment.')
parser.add_argument('--env',
dest='env',
type=str,
help='OpenAI environment',
default="CartPole-v1")
parser.add_argument('--trainer',
dest='trainer_fname',
type=str,
help='Previously saved Trainer object',
default="")
parser.add_argument('--trainer_savename',
dest='trainer_savename',
type=str,
help='Name under which to save the Trainer object',
default="")
parser.add_argument('--agent_savename',
dest='agent_savename',
type=str,
help='Name under which to save the top Agent object',
default="")
parser.add_argument('--training_results_savename',
dest='training_results_savename',
type=str,
help='Name under which to save the training results',
default="")
parser.add_argument('--generations',
dest='num_generations',
type=int,
help='Number of generations',
default=500)
parser.add_argument('--episodes',
dest='num_episodes',
type=int,
help='Number of episodes per agent at each generation',
default=1)
parser.add_argument('--pop',
dest='r_size',
type=int,
help='Number of agents (root teams) per generation',
default=200)
parser.add_argument('--frames',
dest='num_frames',
type=int,
help='Max number of frames per episode',
default=18000)
parser.add_argument('--seed',
dest='seed',
type=int,
help='Seed for environment',
default=-1)
parser.add_argument('--verbose',
dest='verbose',
type=bool,
help="Print results to the console as we evolve",
default=False)
parser.add_argument('--fast',
dest='fast',
type=bool,
help="Set to True to skip re-evaluating agents",
default=False)
parser.add_argument(
'--skips',
dest='skips',
type=int,
help='Maximum number of times an agent can skip re-evaluation',
default=0)
parser.add_argument('--action_type',
dest='action_type',
type=int,
help='0 = Standard, 1 = Real-valued',
default=0)
args = parser.parse_args(arguments)
# Get environment details
env = gym.make(args.env)
if args.action_type == 0:
num_actions = env.action_space.n
elif args.action_type == 1:
num_actions = 100 # This number is meaningless when choosing real-valued actions
# but the Team initialization doesn't like anything less than 2.
# In fact, the bigger the number, the faster Team intialization.
else:
print("Invalid action_type argument {}".format(args.action_type))
input_size = reduce(lambda x, y: x * y, env.observation_space.shape)
configureTrainer(
atomic_action_range=num_actions,
p_del_learner=0.7,
p_add_learner=0.7,
p_mut_learner=0.3,
p_mut_learner_action=0.6,
p_atomic=0.5,
r_size=args.r_size,
percent_keep=0.40,
env_name=args.env,
trainer_name=args.trainer_savename,
agent_name=args.agent_savename,
verbose=args.verbose,
env_seed=args.seed,
do_fast_track=args.fast,
max_eval_skips=args.skips,
action_selection=args.action_type,
training_results_nparray_name=args.training_results_savename)
configureProgram(p_add=0.6,
p_del=0.6,
p_mut=0.6,
p_swap=0.6,
max_prog_size=128,
min_prog_size=8,
num_registers=8,
input_size=input_size)
# If trainer filename was passed in, re-create Trainer. Otherwise, create
# a new one from scratch
if args.trainer_fname == "":
# Create Trainer
trainer = Trainer()
# Perform TPG initialization (ie. create initial Team and Learner populations)
trainer.initialize()
else:
# Load Trainer
trainer = loadTrainer(args.trainer_fname)
# Try to generate an agent
trainer.evolve(env,
num_generations=args.num_generations,
num_episodes=args.num_episodes,
num_frames=args.num_frames)
def run(arguments):
parser = argparse.ArgumentParser(
description='Perform linear GP evolution for a given environment.')
# Program configuration
parser.add_argument('--minps',
dest='min_prog_size',
type=int,
help='Minimum number of instructions per Program',
default=32)
parser.add_argument('--maxps',
dest='max_prog_size',
type=int,
help='Maximum number of instructions per Program',
default=1024)
parser.add_argument('--padd',
dest='padd',
type=float,
help='Instruction addition strength',
default=0.7)
parser.add_argument('--pdel',
dest='pdel',
type=float,
help='Instruction deletion strength',
default=0.7)
parser.add_argument('--pmut',
dest='pmut',
type=float,
help='Instruction mutation strength',
default=0.7)
# Trainer configuration
parser.add_argument(
'--generations',
dest='num_generations',
type=int,
help='Number of generations over which evolution is performed',
default=50)
parser.add_argument('--pop',
dest='population_size',
type=int,
help='Learner population size',
default=200)
parser.add_argument('--keep',
dest='percent_keep',
type=float,
help='Percentage of surviving Learners',
default=0.3)
parser.add_argument('--fast',
dest='fast_mode',
type=bool,
help='Skip some re-evaluations',
default=True)
parser.add_argument(
'--skips',
dest='num_skips',
type=int,
help='Number of generations over which to skip re-evaluation',
default=3)
parser.add_argument(
'--episodes',
dest='num_eps_per_gen',
type=int,
help=
'Number of episodes over which an agent is evaluated each generation',
default=3)
parser.add_argument(
'--verbose',
dest='verbose',
type=bool,
help='Do print out info to the console during evolution',
default=True)
parser.add_argument('--agent',
dest='agent_save_name',
type=str,
help='Name under which to save the evolved agent',
default="")
parser.add_argument('--fitness_sharing',
dest='fitness_sharing',
type=bool,
help='Use fitness sharing to evaluate learners',
default=False)
# Environment configuration
parser.add_argument('--env',
dest='env',
type=str,
help='OpenAI environment',
default="CartPole-v1")
parser.add_argument('--statespace',
dest='statespace',
type=int,
help='Length of flattened state space',
default=4)
args = parser.parse_args(arguments)
# Set env
env, args = set_env(args)
ConfigureProgram(num_inputs=args.statespace,
min_prog_size=args.min_prog_size,
max_prog_size=args.max_prog_size,
p_add=args.padd,
p_del=args.pdel,
p_mut=args.pmut)
ConfigureTrainer(num_generations=args.num_generations,
population_size=args.population_size,
percent_keep=args.percent_keep,
fast_mode=args.fast_mode,
max_num_skips=args.num_skips,
num_eps_per_gen=args.num_eps_per_gen,
verbose=args.verbose,
agent_save_name=args.agent_save_name,
output_folder="../lgp-outputs/",
env_name=args.env,
fitness_sharing=args.fitness_sharing)
trainer = Trainer(env)
trainer.evolve()