def train(topology_id: int,
batch_size: int = 1000,
unused_cpu: int = 2,
use_backup: bool = False):
"""Create an infinite amount of genomes for the requested topology."""
# Get the config
cfg = get_config()
# Get initial genome key
g_key, csv_path = get_initial_keys(topology_id, use_backup=use_backup)
# Setup the environment
_, games = get_game_ids(experiment_id=6)
multi_env = get_multi_env(config=cfg)
multi_env.set_games(games, noise=False)
# Create genomes
t = time.localtime()
print(f"\nCurrent time: {t.tm_hour:02d}h-{t.tm_min:02d}m-{t.tm_sec:02d}s")
print(
f"> Evaluating {batch_size} genomes in csv '{csv_path.split('/')[-1]}'"
)
genomes = list(
iteritems({
g_key + i: get_genome(topology_id, g_id=g_key + i, cfg=cfg)
for i in range(batch_size)
}))
g_key += batch_size
# Evaluate the genome-dictionary in parallel
try:
pool = mp.Pool(mp.cpu_count() - unused_cpu)
manager = mp.Manager()
return_dict = manager.dict()
for genome in genomes:
pool.apply_async(func=multi_env.eval_genome,
args=(genome, return_dict))
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
# Calculate the fitness from the given return_dict
fitness = calc_pop_fitness(
fitness_cfg=cfg.evaluation,
game_cfg=cfg.game,
game_obs=return_dict,
)
for i, genome in genomes:
genome.fitness = fitness[i]
# Write the result to CSV
with open(csv_path, 'a', newline='') as f:
writer = csv.writer(f)
for _, g in genomes:
writer.writerow(
get_genome_parameters(g, topology_id=topology_id))
except KeyboardInterrupt:
# Remove the temporary CSV first
os.remove(csv_path)
raise KeyboardInterrupt
def evaluate_generations(name,
experiment_id,
folder=None,
hops: int = 10,
unused_cpu: int = 2):
"""
Evaluate the population across its lifetime. At each generation, the ten best genomes are evaluated together with
the elite genome of the past five generations.
:param name: Name of the population
:param experiment_id: Experiment for which the population is trained (and now will be evaluated)
:param folder: Population-folder (~experiment level)
:param hops: Number of generations between each saved population
:param unused_cpu: Number of CPU cores not used
"""
# Fetch population and evaluation games
folder = folder if folder else get_folder(experiment_id)
pop = Population(
name=name,
folder_name=folder,
log_print=False,
use_backup=True,
)
_, game_ids_eval = get_game_ids(experiment_id=experiment_id)
# Perform the evaluations
max_gen = pop.generation
for gen in tqdm(range(0, max_gen + 1, hops)):
# Load in the current generation
if not pop.load(gen=gen):
raise Exception(
f"Population {name} is not trained for generation {gen}")
# Collect the used genomes
if gen > 5:
genomes = sorted([g for g in pop.population.values()],
key=lambda x: x.fitness if x.fitness else 0,
reverse=True)[:10]
for i in range(1, 6):
keys = [g.key for g in genomes]
g = copy.deepcopy(pop.best_genome_hist[gen - i]
[1]) # Copy since chance of mutation
while g.key in keys: # Already added to genomes, update keys
g.key += 1
genomes.append(g)
else:
# No history yet, use only the ten most fit genomes from the current generation
genomes = sorted([g for g in pop.population.values()],
key=lambda x: x.fitness if x.fitness else 0,
reverse=True)[:15]
# Evaluate the selected genomes
evaluate(
population=pop,
games=game_ids_eval,
genomes=genomes,
unused_cpu=unused_cpu,
overwrite=True,
)
def evaluate_population(pop: Population, cfg: Config, cpu: int, experiment_id: int):
"""Evaluate the given population."""
pop.log(f"{pop.name} - Evaluating the population...")
_, game_ids_eval = get_game_ids(experiment_id=experiment_id)
multi_env = get_multi_env(pop=pop, game_config=cfg)
multi_env.set_games(game_ids_eval, noise=False)
pool = mp.Pool(mp.cpu_count() - cpu)
manager = mp.Manager()
return_dict = manager.dict()
pbar = tqdm(total=len(pop.population), desc="Evaluating")
def update(*_):
pbar.update()
for genome_id, genome in pop.population.items():
pool.apply_async(func=multi_env.eval_genome, args=((genome_id, genome), return_dict), callback=update)
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
pbar.close()
# Calculate the fitness from the given return_dict
pop.log(f"{pop.name} - Calculating fitness scores...")
fitness = calc_pop_fitness(
fitness_cfg=pop.config.evaluation,
game_cfg=cfg.game,
game_obs=return_dict,
gen=pop.generation,
)
for i, genome in pop.population.items():
genome.fitness = fitness[i]
# Get the fittest genome
best = None
for g in pop.population.values():
if best is None or g.fitness > best.fitness: best = g
pop.best_genome = best
# Save the results
pop.save()
# Visualize most fit genome
visualize_genome(
debug=True,
genome=best,
population=pop,
)
# Trace the most fit genome
trace_most_fit(
debug=False,
games=game_ids_eval,
genome=best,
population=pop,
unused_cpu=cpu,
)
def simulation(cython: bool = False, parallel: bool = False):
"""Execute a simulation of one training iteration."""
# Create the population
cfg = Config()
cfg.game.duration = 200
cfg.update()
pop = Population(
name="delete_me",
config=cfg,
folder_name="test_performance",
use_backup=True,
overwrite=True, # Every iteration, create a new population from scratch
)
# Perform the simulations of experiment3
train_games, _ = get_game_ids(experiment_id=3)
if cython:
multi_env = MultiEnvironmentCy(game_config=cfg, pop_config=pop.config)
else:
multi_env = MultiEnvironment(game_config=cfg, pop_config=pop.config)
multi_env.set_games(train_games, noise=True)
# Prepare the generation's reporters for the generation
pop.reporters.start_generation(gen=pop.generation, logger=pop.log)
# Fetch the dictionary of genomes
genomes = list(iteritems(pop.population))
if parallel:
# Initialize the evaluation-pool
pool = mp.Pool(mp.cpu_count())
manager = mp.Manager()
return_dict = manager.dict()
for genome in genomes:
pool.apply_async(func=multi_env.eval_genome,
args=(genome, return_dict))
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
else:
return_dict = dict()
for genome in genomes:
multi_env.eval_genome(genome, return_dict)
path = f"population_backup/storage/test_performance/"
if os.path.exists(path):
shutil.rmtree(path)
# Get game's blueprint
game.get_blueprint(show_player=True)
plt.title("Blueprint - Game {id:05d}".format(id=game.id))
plt.savefig(f'environment/visualizations/blueprint_game{game.id:05d}')
if show: plt.show()
plt.close()
if __name__ == '__main__':
os.chdir("../..")
cfg = Config()
games = [0] # Game Dummy
# Experiment 1
exp1_train, exp1_eval = get_game_ids(1)
games += exp1_train[:1] # Repeats itself
games += exp1_eval
# Experiment 2
exp2_train, exp2_eval = get_game_ids(2)
games += exp2_train[:1] # Repeats itself
games += exp2_eval
# Experiment 3
exp3_train, exp3_eval = get_game_ids(3)
games += exp3_train[:1] # Repeats itself
games += exp3_eval
# Experiment 6
exp6_train, exp6_eval = get_game_ids(6)
def main(
fitness,
prob_gru: float,
prob_gru_nr: float,
prob_gru_nu: float,
prob_simple_rnn: float,
train_iterations=0,
version=0,
unused_cpu=1,
):
"""
Run a population's configuration.
:param fitness: Fitness function used to evaluate the population
:param prob_gru: Probability of mutating towards a GRU-node
:param prob_gru_nr: Probability of mutating towards a GRU-NR-node
:param prob_gru_nu: Probability of mutating towards a GRU-NU-node
:param prob_simple_rnn: Probability of mutating towards a SimpleRNN-node
:param train_iterations: Number of training generations
:param version: Version of the model
:param unused_cpu: Number of CPUs not used during training
"""
# Re-configure the config-file
cfg = Config()
cfg.bot.angular_dir = []
cfg.bot.delta_dist_enabled = False
cfg.bot.dist_enabled = True
cfg.game.duration = 60 # 60 seconds should be enough to reach the target from each starting orientation
cfg.genome.node_add_prob = 0 # Do not change number of hidden nodes
cfg.genome.node_disable_prob = 0 # Do not change number of hidden nodes
cfg.population.pop_size = 512
cfg.population.compatibility_thr = 1. # Very small since all architectures have strictly one hidden node
# Let inputs apply to configuration
cfg.genome.rnn_prob_gru = prob_gru
cfg.genome.rnn_prob_gru_nr = prob_gru_nr
cfg.genome.rnn_prob_gru_nu = prob_gru_nu
cfg.genome.rnn_prob_simple_rnn = prob_simple_rnn
cfg.evaluation.fitness = fitness
cfg.update()
# Create the population
folder = get_folder(experiment_id=4)
name = get_name(cfg=cfg, version=version)
pop = Population(
name=name,
config=cfg,
folder_name=folder,
use_backup=False,
)
# Make sure that all of the genomes in the initial population have exactly one hidden node
if pop.generation == 0:
for g in pop.population.values():
g.mutate_add_node(pop.config.genome)
# Give overview of population
gru = cfg.genome.rnn_prob_gru
gru_nr = cfg.genome.rnn_prob_gru_nr
gru_nu = cfg.genome.rnn_prob_gru_nu
rnn = cfg.genome.rnn_prob_simple_rnn
msg = f"\n\n\n\n\n===> RUNNING EXPERIMENT 4 FOR THE FOLLOWING CONFIGURATION: <===" \
f"\n\t> fitness: {cfg.evaluation.fitness}" \
f"\n\t> GRU enabled: {gru > 0} (probability={round(gru, 2)})" \
f"\n\t> GRU-NR enabled: {gru_nr > 0} (probability={round(gru_nr, 2)})" \
f"\n\t> GRU-NU enabled: {gru_nu > 0} (probability={round(gru_nu, 2)})" \
f"\n\t> SRU enabled: {rnn > 0} (probability={round(rnn, 2)})" \
f"\n\t> Saving under folder: {folder}" \
f"\n\t> Training iterations: {train_iterations}\n"
pop.log(msg)
# Set games used for evaluation
games_train, games_eval = get_game_ids(experiment_id=4)
# Execute the requested segments
try:
train(
game_config=cfg,
games=games_train,
iterations=train_iterations,
population=pop,
unused_cpu=unused_cpu,
)
# Evaluate the trained population
evaluate(
games=games_eval,
population=pop,
unused_cpu=unused_cpu,
)
training_overview(population=pop, )
visualize_genome(
genome=pop.best_genome,
population=pop,
)
# Perform GRU-analysis if population is NEAT-GRU
if gru > 0:
gru_analysis(
population=pop,
unused_cpu=unused_cpu,
experiment_id=4,
)
monitor(
game_id=games_eval[0],
population=pop,
genome=pop.best_genome,
)
except Exception as e:
pop.log(traceback.format_exc(), print_result=False)
raise e
finally:
process_killer('run_population.py') # Close all the terminated files
plt.imshow(result)
plt.tight_layout()
plt.savefig(f"{path[:-1]}.png", bbox_inches='tight', pad_inches=0)
plt.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='')
parser.add_argument('--game_id', type=int, default=10002)
parser.add_argument('--debug', type=bool, default=False)
args = parser.parse_args()
# Go back to root
os.chdir("../../../")
_, eval_games = get_game_ids(experiment_id=4)
# Create the population
config = Config()
pop = Population(
name='test',
# name='NEAT-SRU/v1',
folder_name='test',
# folder_name=get_folder(args.experiment),
config=config,
)
chosen_genome = pop.best_genome
# chosen_genome = pop.population[47280]
# Run the monitor-tool
main(
def main(
fitness,
prob_gru: float,
prob_sru: float,
prob_lstm: float,
version=0,
unused_cpu=1,
):
"""
Run a population's configuration.
:param fitness: Fitness function used to evaluate the population
:param prob_gru: Probability of mutating towards a GRU-node
:param prob_sru: Probability of mutating towards a SRU-node
:param prob_lstm: Probability of mutating towards a LSTM-node
:param version: Version of the model
:param unused_cpu: Number of CPUs not used during training
"""
# Re-configure the config-file
cfg = Config()
cfg.bot.angular_dir = []
cfg.bot.delta_dist_enabled = False
cfg.bot.dist_enabled = True
cfg.game.duration = 60 # 60 seconds should be enough to reach the target from each starting orientation
cfg.population.pop_size = 512
# Let inputs apply to configuration
cfg.genome.rnn_prob_gru = prob_gru
cfg.genome.rnn_prob_simple_rnn = prob_sru
cfg.genome.rnn_prob_lstm = prob_lstm
cfg.evaluation.fitness = fitness
cfg.update()
# Copy population over from experiment1
name = get_name(cfg=cfg, version=version)
path_exp1 = f'population/storage/experiment1/{name}/'
if not os.path.exists(path_exp1):
raise Exception(
f"Experiment 1 must be executed first for population {name}, terminating experiment 2..."
)
# Population exists in experiment1, copy over to experiment2 (change experiment1 population's folder and save)
pop = Population(name=name,
config=cfg,
folder_name=get_folder(experiment_id=1),
use_backup=False)
assert pop.generation > 0 # Population is not new (redundant check)
folder = get_folder(experiment_id=2)
pop.folder_name = folder
pop.save() # Overrides pre-existing populations!
# Copy over all generations as well, since these are used during population evaluation
path = f"population{'_backup' if pop.use_backup else ''}/storage/{pop.folder_name}/{pop}/"
copy_tree(f"{path_exp1}generations", f"{path}generations")
# Give overview of population
gru = cfg.genome.rnn_prob_gru
sru = cfg.genome.rnn_prob_simple_rnn
lstm = cfg.genome.rnn_prob_lstm
msg = f"\n\n\n\n\n===> RUNNING EXPERIMENT 2 FOR THE FOLLOWING CONFIGURATION: <===" \
f"\n\t> fitness: {cfg.evaluation.fitness}" \
f"\n\t> GRU enabled: {gru > 0} (probability={round(gru, 2)})" \
f"\n\t> SRU enabled: {sru > 0} (probability={round(sru, 2)})" \
f"\n\t> LSTM enabled: {lstm > 0} (probability={round(lstm, 2)})" \
f"\n\t> Saving under folder: {folder}\n"
pop.log(msg)
# Set games used for evaluation
_, games_eval = get_game_ids(experiment_id=2)
# Execute the requested segments
try:
# Evaluate the trained population
evaluate(
games=games_eval,
population=pop,
unused_cpu=unused_cpu,
)
except Exception as e:
pop.log(traceback.format_exc(), print_result=False)
raise e
finally:
process_killer('run_population.py') # Close all the terminated files
def main(pop_name: str,
version: int,
unused_cpu: int = 2,
use_backup: bool = False):
# Check if valid population name
if pop_name not in SUPPORTED:
raise Exception(f"Population '{pop_name}' not supported!")
# Create the population
cfg = get_config()
cfg.population.specie_elitism = 1
folder = get_folder(experiment_id=7)
pop = Population(
name=f'{pop_name}/v{version}',
config=cfg,
folder_name=folder,
use_backup=use_backup,
)
# Replace the population's initial population with the requested topologies genomes
if pop.generation == 0:
for g_id in pop.population.keys():
pop.population[g_id] = get_topology(pop_name, gid=g_id, cfg=cfg)
pop.species.speciate(config=pop.config,
population=pop.population,
generation=pop.generation,
logger=pop.log)
pop.log(f"\n\n\n===> RUNNING EXPERIMENT 7 <===\n")
# Set games and environment used for training and evaluation
games_train, games_eval = get_game_ids(experiment_id=7)
train_env = get_multi_env(config=cfg)
eval_env = get_multi_env(config=cfg)
eval_env.set_games(games_eval, noise=False)
solution_found = False
while not solution_found:
# Train the population for a single iteration
pop.log("\n===> TRAINING <===")
train_env.set_games(games_train, noise=True)
# Prepare the generation's reporters for the generation
pop.reporters.start_generation(gen=pop.generation, logger=pop.log)
# Fetch the dictionary of genomes
genomes = list(iteritems(pop.population))
# Initialize the evaluation-pool
pool = mp.Pool(mp.cpu_count() - unused_cpu)
manager = mp.Manager()
return_dict = manager.dict()
for genome in genomes:
pool.apply_async(func=train_env.eval_genome,
args=(genome, return_dict))
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
# Calculate the fitness from the given return_dict
fitness = calc_pop_fitness(
fitness_cfg=pop.config.evaluation,
game_cfg=cfg.game,
game_obs=return_dict,
gen=pop.generation,
)
for i, genome in genomes:
genome.fitness = fitness[i]
# Update the population's best_genome
best = None
for g in itervalues(pop.population):
if best is None or g.fitness > best.fitness: best = g
pop.reporters.post_evaluate(population=pop.population,
species=pop.species,
best_genome=best,
logger=pop.log)
# Update the population's best_genome
genomes = sorted(pop.population.items(),
key=lambda x: x[1].fitness,
reverse=True)
pop.best_fitness[pop.generation] = genomes[0][1].fitness
pop.best_genome_hist[pop.generation] = genomes[0]
pop.best_genome = best
# Let population evolve
evolve(pop, pop_name)
# End generation
pop.reporters.end_generation(population=pop.population,
name=str(pop),
species_set=pop.species,
logger=pop.log)
# Test if evaluation finds a solution for the new generation, impossible if fitness < 0.7
if pop.best_genome.fitness > 0.7 or pop.generation % 10 == 0:
pop.log("\n===> EVALUATING <===")
genomes = list(iteritems(pop.population))
pool = mp.Pool(mp.cpu_count() - unused_cpu)
manager = mp.Manager()
return_dict = manager.dict()
for genome in genomes:
pool.apply_async(func=eval_env.eval_genome,
args=(genome, return_dict))
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
# Calculate the fitness from the given return_dict
finished = calc_finished_ratio(
fitness_cfg=cfg.evaluation,
game_obs=return_dict,
)
best = None
for i, genome in genomes:
genome.fitness = finished[i]
if best is None or finished[i] > best.fitness: best = genome
pop.log(f"Best genome:\n{best}\n{best.nodes[2]}")
# Solution is found
if best.fitness == 1:
pop.best_genome = best
pop.log(f"Solution found!")
solution_found = True # End the outer while-loop
# Save the population with their evaluation results
pop.save()
def main(topology_id: int,
iterations: int,
eval_interval: int,
experiment_id: int,
hops: float = 0.1,
weight_range: float = 1,
mutate_combine: bool = False,
mutate_bias: bool = True,
mutate_reset: bool = True,
mutate_update: bool = True,
mutate_candidate: bool = True,
init_population_size: int = 1000,
unused_cpu: int = 2):
"""
Run the fifth experiment.
:param topology_id: Chosen topology to investigate
:param iterations: Number of training iterations performed
:param eval_interval: After how much training iterations evaluation is performed
:param experiment_id: ID of the experiment used to train and evaluate the population on
:param hops: Hops between variable-configurations
:param weight_range: Range of deviation for one's weights
:param mutate_combine: Combine the best results of each mutation type
:param mutate_bias: Mutate the GRU's bias values
:param mutate_reset: Mutate the reset-gate's weights
:param mutate_update: Mutate the update-gate's weights
:param mutate_candidate: Mutate the candidate-state's weights
:param init_population_size: Initial size of the randomized population
:param unused_cpu: Number of CPU-cores not used
"""
# Get the population
name = f"experiment{experiment_id}_topology{topology_id}_hops{hops}_range{weight_range}"
pop = Population(
name=name,
folder_name='experiment5',
use_backup=False,
)
# Define the games specific to the experiment
_, game_ids_eval = get_game_ids(experiment_id=experiment_id)
# Set the genomes if population is new
if pop.generation == 0:
# Get the requested topology-type
if topology_id == 1:
topology = get_topology1
elif topology_id == 2:
topology = get_topology2
else:
raise Exception(f"Topology {topology_id} not supported.")
# Initialize the population with a randomized population
pop.population = dict()
for gid in range(init_population_size):
new_genome = topology(pop.config, random_init=True)
new_genome.key = gid
pop.population[gid] = new_genome
# Perform an initial training
train_population(pop=pop, games=game_ids_eval, unused_cpu=unused_cpu)
pop.generation = 0 # Don't count initial training as a generation
# Get the fittest genome
best = None
for g in pop.population.values():
if best is None or g.fitness > best.fitness: best = g
pop.best_genome = best
visualize_best_genome(pop=pop)
# Test the initial population
test_population(pop=pop, games=game_ids_eval)
# Set the most fit genome as the starting-point
set_population(pop=pop,
hops=hops,
weight_range=weight_range,
mutate_bias=mutate_bias,
mutate_reset=mutate_reset,
mutate_update=mutate_update,
mutate_candidate=mutate_candidate)
# Evaluate the population
for i in range(iterations):
train_population(pop=pop, games=game_ids_eval, unused_cpu=unused_cpu)
evaluate_fitness(pop=pop,
hops=hops,
weight_range=weight_range,
mutate_combine=mutate_combine,
mutate_bias=mutate_bias,
mutate_reset=mutate_reset,
mutate_update=mutate_update,
mutate_candidate=mutate_candidate)
visualize_best_genome(pop=pop)
if pop.generation % eval_interval == 0:
test_population(pop=pop, games=game_ids_eval)
set_population(pop=pop,
hops=hops,
weight_range=weight_range,
mutate_bias=mutate_bias,
mutate_reset=mutate_reset,
mutate_update=mutate_update,
mutate_candidate=mutate_candidate)
def main(topology_id: int,
batch_size: int = 1000,
train_batch: int = 3,
min_finished: float = MIN_FINISHED,
unused_cpu: int = 2,
save_pop: bool = False,
use_backup: bool = False):
"""Run a population infinitely long and store all its good genomes."""
# Get the CSV used to store the results in
csv_path, csv_name, added = get_csv_path(topology_id, use_backup=use_backup, batch_size=batch_size)
# Create the population
name = csv_name if save_pop else 'dummy'
cfg = get_config()
folder = get_folder(experiment_id=6)
pop = Population(
name=name,
config=cfg,
folder_name=folder,
use_backup=use_backup,
overwrite=True, # Every iteration, create a new population from scratch
)
# Replace the population's initial population with the requested topologies genomes
for g_id in pop.population.keys(): pop.population[g_id] = get_genome(topology_id, g_id=g_id, cfg=cfg)
pop.species.speciate(config=pop.config,
population=pop.population,
generation=pop.generation,
logger=pop.log)
# Set games and environment used for training and evaluation
pop.log(f"\n\n\n===> RUNNING EXPERIMENT 6 <===\n")
games_train, games_eval = get_game_ids(experiment_id=6)
train_env = get_multi_env(config=cfg)
eval_env = get_multi_env(config=cfg)
eval_env.set_games(games_eval, noise=False)
# Keep training and evolving the network until the complete CSV is filled
last_saved = pop.generation
try:
while added < batch_size:
t = time.localtime()
pop.log(f"\n\n===> Selective genome creation at {added / batch_size * 100}%, "
f"storing in csv '{csv_path.split('/')[-1]}' "
f"({t.tm_hour:02d}h-{t.tm_min:02d}m-{t.tm_sec:02d}s) <===")
# Train the population
pop.log("\n===> Training <===")
for _ in tqdm(range(train_batch), desc="Training"):
train_env.set_games(games_train, noise=True)
genomes = list(iteritems(pop.population))
# Initialize the evaluation-pool
pool = mp.Pool(mp.cpu_count() - unused_cpu)
manager = mp.Manager()
return_dict = manager.dict()
for genome in genomes:
pool.apply_async(func=train_env.eval_genome, args=(genome, return_dict))
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
# Calculate the fitness from the given return_dict
fitness = calc_pop_fitness(
fitness_cfg=pop.config.evaluation,
game_cfg=cfg.game,
game_obs=return_dict,
gen=pop.generation,
)
for i, genome in genomes:
genome.fitness = fitness[i]
# Update the population's best_genome
best = None
for g in itervalues(pop.population):
if best is None or g.fitness > best.fitness: best = g
genomes = sorted(pop.population.items(), key=lambda x: x[1].fitness, reverse=True)
pop.best_fitness[pop.generation] = genomes[0][1].fitness
pop.best_genome_hist[pop.generation] = genomes[0]
pop.best_genome = best
pop.log(f"Best training fitness: {best.fitness}")
# Let population evolve
pop.evolve()
# Constraint each of the population's new genomes to the given topology
for g in pop.population.values():
enforce_topology(g, topology_id=topology_id)
# Save the population after training
if pop.generation - last_saved >= 100:
pop.save()
last_saved = pop.generation
# Evaluate the current population as was done in experiment6
pop.log("\n===> EVALUATING <===")
genomes = list(iteritems(pop.population))
pool = mp.Pool(mp.cpu_count() - unused_cpu)
manager = mp.Manager()
return_dict = manager.dict()
for genome in genomes:
pool.apply_async(func=eval_env.eval_genome, args=(genome, return_dict))
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
# Calculate the fitness from the given return_dict
finished = calc_finished_ratio(
fitness_cfg=cfg.evaluation,
game_obs=return_dict,
)
best = None
for i, genome in genomes:
genome.fitness = finished[i]
if best is None or finished[i] > best.fitness: best = genome
# Give evaluation overview of population
pop.log(f"Best evaluation finish ratio: {round(best.fitness, 2)}")
best_str = str(best).replace("\n", "\n\t")
best_str += "\n\t" + str(best.nodes[2]).replace("\n", "\n\t")
pop.log(f"Best genome: \n\t{best_str}")
sids = list(iterkeys(pop.species.species))
sids.sort()
msg = f"\nPopulation '{name}' has {len(pop.species.species):d} species:" \
f"\n\t specie age size finished stag " \
f"\n\t======== ===== ====== ========== ======"
pop.log(msg) if pop.log else print(msg)
for sid in sids:
s = pop.species.species[sid]
a = pop.generation - s.created
n = len(s.members)
sf = [g.fitness for g in s.members.values() if g.fitness]
f = "--" if len(sf) == 0 else f"{max(sf):.2f}"
st = pop.generation - s.last_improved
msg = f"\t{sid:^8} {a:^5} {n:^6} {f:^10} {st:^6}"
pop.log(msg) if pop.log else print(msg)
# Write the result to CSV
with open(csv_path, 'a', newline='') as f:
writer = csv.writer(f)
for _, g in genomes:
# Only write the genomes that exceed the minimum 'finished ratio' threshold!
if g.fitness >= min_finished:
writer.writerow(get_genome_parameters(g, topology_id=topology_id))
added += 1
finally:
# Remove the dummy population if it exists
pop.save()
path = f"population{'_backup' if use_backup else ''}/storage/{pop.folder_name}/dummy/"
if os.path.exists(path):
shutil.rmtree(path)
def main(
fitness,
prob_gru: float,
prob_sru: float,
prob_lstm: float,
train_iterations=0,
version=0,
unused_cpu=1,
):
"""
Run a population's configuration.
:param fitness: Fitness function used to evaluate the population
:param prob_gru: Probability of mutating towards a GRU-node
:param prob_sru: Probability of mutating towards a SRU-node
:param prob_lstm: Probability of mutating towards a LSTM-node
:param train_iterations: Number of training generations
:param version: Version of the model
:param unused_cpu: Number of CPUs not used during training
"""
# Re-configure the config-file
cfg = Config()
cfg.bot.angular_dir = []
cfg.bot.delta_dist_enabled = False
cfg.bot.dist_enabled = True
cfg.game.duration = 60 # 60 seconds should be enough to reach the target from each starting orientation
cfg.population.pop_size = 512
# Let inputs apply to configuration
cfg.genome.rnn_prob_gru = prob_gru
cfg.genome.rnn_prob_simple_rnn = prob_sru
cfg.genome.rnn_prob_lstm = prob_lstm
cfg.evaluation.fitness = fitness
cfg.update()
# Create the population
folder = get_folder(experiment_id=1)
name = get_name(cfg=cfg, version=version)
pop = Population(
name=name,
config=cfg,
folder_name=folder,
use_backup=False,
)
# Give overview of population
gru = cfg.genome.rnn_prob_gru
sru = cfg.genome.rnn_prob_simple_rnn
lstm = cfg.genome.rnn_prob_lstm
msg = f"\n\n\n\n\n===> RUNNING EXPERIMENT 1 FOR THE FOLLOWING CONFIGURATION: <===" \
f"\n\t> fitness: {cfg.evaluation.fitness}" \
f"\n\t> GRU enabled: {gru > 0} (probability={round(gru, 2)})" \
f"\n\t> SRU enabled: {sru > 0} (probability={round(sru, 2)})" \
f"\n\t> LSTM enabled: {lstm > 0} (probability={round(lstm, 2)})" \
f"\n\t> Saving under folder: {folder}" \
f"\n\t> Training iterations: {train_iterations}\n"
pop.log(msg)
# Set games used for evaluation
games_train, games_eval = get_game_ids(experiment_id=1)
# Execute the requested segments
try:
train(
debug=False,
games=games_train,
iterations=train_iterations,
population=pop,
unused_cpu=unused_cpu,
)
# Evaluate the trained population
evaluate(
games=games_eval,
population=pop,
unused_cpu=unused_cpu,
)
training_overview(population=pop, )
visualize_genome(
genome=pop.best_genome,
population=pop,
)
except Exception as e:
pop.log(traceback.format_exc(), print_result=False)
raise e
finally:
process_killer('run_population.py') # Close all the terminated files
