def train_population(pop: Population, games: list, unused_cpu: int = 2):
"""Evaluate the given population on a training set."""
multi_env = get_multi_env(pop=pop, game_config=pop.config)
multi_env.set_games(games, noise=False)
pool = mp.Pool(mp.cpu_count() - unused_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
fitness = calc_pop_fitness(
fitness_cfg=pop.config.evaluation,
game_cfg=pop.config.game,
game_obs=return_dict,
gen=pop.generation,
)
for i, genome in pop.population.items():
genome.fitness = fitness[i]
# Save the results
pop.generation += 1
pop.save()
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_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 evaluate_genome_list(self,
genome_list,
pop: Population,
parallel: bool = True,
overwrite: bool = False):
"""
Evaluate the population for a single evaluation-process.
:param genome_list: List of genomes that will be evaluated
:param pop: The population to which the genomes belong (used to setup the network and query the config)
:param parallel: Evaluate the given genomes in parallel
:param overwrite: Overwrite the evaluation dictionary if it already exists
"""
# Create the environment which is responsible for evaluating the genomes
multi_env = get_multi_env(pop=pop, game_config=self.game_config)
# Evaluate on all the games
multi_env.set_games(self.games, noise=False)
# Fetch requested genomes
genomes = [(g.key, g) for g in genome_list]
if parallel:
pool = mp.Pool(mp.cpu_count() - self.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
else:
return_dict = dict()
for genome in tqdm(genomes, desc="sequential evaluating"):
multi_env.eval_genome(genome, return_dict)
# Create the evaluation for each of the genomes
eval_result = dict()
for k in return_dict.keys():
# Create answer based on game.close()
eval_result[str(k)] = create_answer(return_dict[k])
# Append fitness to answer
eval_result[str(k)][D_FITNESS] = calc_pop_fitness(
fitness_cfg=pop.config.evaluation,
game_cfg=pop.config.game,
game_obs={k: return_dict[k]},
gen=pop.generation)[k]
pop.add_evaluation_result(eval_result, overwrite=overwrite)
def train(
population: Population,
game_config: Config,
games: list,
iterations: int,
unused_cpu: int = 0,
save_interval: int = 10,
):
"""Train the population on the requested number of iterations. Manual adaptation of main's train()."""
population.log("\n===> TRAINING <===\n")
multi_env = get_multi_env(pop=population, game_config=game_config)
msg = f"Repetitive evaluating on games: {games} for {iterations} iterations"
population.log(msg, print_result=False)
# Iterate and evaluate over the games
saved = True
for iteration in range(iterations):
# Set and randomize the games
multi_env.set_games(games, noise=True)
# Prepare the generation's reporters for the generation
population.reporters.start_generation(gen=population.generation,
logger=population.log)
# Fetch the dictionary of genomes
genomes = list(iteritems(population.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=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=population.config.evaluation,
game_cfg=game_config.game,
game_obs=return_dict,
gen=population.generation,
)
for i, genome in genomes:
genome.fitness = fitness[i]
# Gather and report statistics
best = None
for g in itervalues(population.population):
if best is None or g.fitness > best.fitness: best = g
population.reporters.post_evaluate(population=population.population,
species=population.species,
best_genome=best,
logger=population.log)
# Update the population's best_genome
genomes = sorted(population.population.items(),
key=lambda x: x[1].fitness,
reverse=True)
population.best_fitness[population.generation] = genomes[0][1].fitness
population.best_genome_hist[population.generation] = genomes[0]
population.best_genome = best
# Let population evolve
population.evolve()
# Update the genomes such all have one hidden node
for g in population.population.values():
n_hidden, _ = g.size()
while n_hidden < 1:
g.mutate_add_connection(population.config.genome)
n_hidden, _ = g.size()
# End generation
population.reporters.end_generation(population=population.population,
name=str(population),
species_set=population.species,
logger=population.log)
# Save the population
if (iteration + 1) % save_interval == 0:
population.save()
saved = True
else:
saved = False
# Make sure that last iterations saves
if not saved: population.save()
def single_evaluation(multi_env, parallel: bool, pop: Population,
unused_cpu: int):
"""
Perform a single evaluation-iteration.
:param multi_env: Environment used to execute the game-simulation in
:param parallel: Boolean indicating if training happens in parallel or not
:param pop: Population used to evaluate on
:param unused_cpu: Number of CPU-cores not used during evaluation
"""
# 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:
pbar = tqdm(total=len(genomes), desc="parallel training")
# Initialize the evaluation-pool
pool = mp.Pool(mp.cpu_count() - unused_cpu)
manager = mp.Manager()
return_dict = manager.dict()
def cb(*_):
"""Update progressbar after finishing a single genome's evaluation."""
pbar.update()
for genome in genomes:
pool.apply_async(func=multi_env.eval_genome,
args=(genome, return_dict),
callback=cb)
pool.close() # Close the pool
pool.join() # Postpone continuation until everything is finished
pbar.close() # Close the progressbar
else:
return_dict = dict()
for genome in tqdm(genomes, desc="sequential training"):
multi_env.eval_genome(genome, return_dict)
# Calculate the fitness from the given return_dict
try:
fitness = calc_pop_fitness(
fitness_config=pop.config.evaluation,
game_observations=return_dict,
game_params=multi_env.get_game_params(),
generation=pop.generation,
)
for i, genome in genomes:
genome.fitness = fitness[i]
except Exception: # TODO: Fix! Sometimes KeyError in fitness (path problem)
pop.log(
f"Exception at fitness calculation: \n{traceback.format_exc()}",
print_result=False)
warnings.warn(
f"Exception at fitness calculation: \n{traceback.format_exc()}")
# Set fitness to zero for genomes that have no fitness set yet
for i, genome in genomes:
if not genome.fitness: genome.fitness = 0.0
# Gather and report statistics
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_genome_hist[
pop.generation] = genomes[:pop.config.population.genome_elitism]
if pop.best_genome is None or best.fitness > pop.best_genome.fitness:
pop.best_genome = best
# Let population evolve
pop.evolve()
# End generation
pop.reporters.end_generation(population=pop.population,
name=str(pop),
species_set=pop.species,
logger=pop.log)
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 evaluate_and_evolve(
self,
pop: Population,
n: int = 1,
parallel=True,
save_interval: int = 1,
):
"""
Evaluate the population on the same set of games.
:param pop: Population object
:param n: Number of generations
:param parallel: Parallel the code (disable parallelization for debugging purposes)
:param save_interval: Indicates how often a population gets saved
"""
multi_env = get_multi_env(pop=pop, game_config=self.game_config)
msg = f"Repetitive evaluating on games: {self.games} for {n} iterations"
pop.log(msg, print_result=False)
# Iterate and evaluate over the games
saved = True
for iteration in range(n):
# Set and randomize the games
multi_env.set_games(self.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() - self.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
else:
return_dict = dict()
for genome in tqdm(genomes, desc="sequential training"):
multi_env.eval_genome(genome, return_dict)
# Calculate the fitness from the given return_dict
fitness = calc_pop_fitness(
fitness_cfg=pop.config.evaluation,
game_cfg=self.game_config.game,
game_obs=return_dict,
gen=pop.generation,
)
for i, genome in genomes:
genome.fitness = fitness[i]
# Gather and report statistics
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
pop.evolve()
# End generation
pop.reporters.end_generation(population=pop.population,
name=str(pop),
species_set=pop.species,
logger=pop.log)
# Save the population
if (iteration + 1) % save_interval == 0:
pop.save()
saved = True
else:
saved = False
# Make sure that last iterations saves
if not saved: pop.save()
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)
