def start_cycle(self, one_by_one=False):
"""Start the cycle.
Keyword arguments:
one_by_one -- evaluate the genomes one by one if True (default False)
"""
if one_by_one:
# play each genome in a game alone.
for generation in range(self.generations):
# retrieve genome list and call evaluation function for each one.
genome_list = NEAT.GetGenomeList(self.population)
best_fitness = 0
print("generation", generation + 1, ":")
print("testing " + str(self.params.PopulationSize) +
" genomes : ")
i = 0
for genome in genome_list:
i += 1
print(i, end=" ")
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
fitness = self.evaluate(genome, generation + 1, i)
if best_fitness < fitness:
best_fitness = fitness
genome.SetFitness(fitness)
# print best fitness and advance to the next generation
print("best fitness : ", best_fitness)
print("=======================================")
self.population.Epoch()
else:
# play all of the population at the same time
for generation in range(self.generations):
# retrieve genome list and build the players list.
genome_list = NEAT.GetGenomeList(self.population)
players = list()
for genome in genome_list:
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
players.append(Dino_player_neat(net))
# start game and retrieve fitness list.
the_game = Dino_NEAT(players, generation)
fitness = the_game.on_execute()
if fitness is None:
print("Training stopped.")
break
# assign each genome to its corresponding fitness.
best_fitness = 0
for i in range(len(fitness)):
genome = genome_list[i]
if best_fitness < fitness[i]:
best_fitness = fitness[i]
genome.SetFitness(fitness[i])
# print best fitness and advance to the next generation
print("generation", generation, ":", best_fitness)
self.population.Epoch()
def getbest(run):
g = NEAT.Genome(0, substrate.GetMinCPPNInputs(), 0,
substrate.GetMinCPPNOutputs(), False,
NEAT.ActivationFunction.TANH, NEAT.ActivationFunction.TANH,
0, params)
pop = NEAT.Population(g, params, True, 1.0, run)
for generation in range(1000):
# Evaluate genomes
genome_list = NEAT.GetGenomeList(pop)
fitnesses = EvaluateGenomeList_Serial(genome_list,
evaluate_xor,
display=False)
[
genome.SetFitness(fitness)
for genome, fitness in zip(genome_list, fitnesses)
]
print('Gen: %d Best: %3.5f' % (generation, max(fitnesses)))
# Print best fitness
# print("---------------------------")
# print("Generation: ", generation)
# print("max ", max([x.GetLeader().GetFitness() for x in pop.Species]))
# Visualize best network's Genome
net = NEAT.NeuralNetwork()
pop.Species[0].GetLeader().BuildPhenotype(net)
img = np.zeros((500, 500, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 500, 500), net)
cv2.imshow("CPPN", img)
# Visualize best network's Pheotype
net = NEAT.NeuralNetwork()
pop.Species[0].GetLeader().BuildESHyperNEATPhenotype(
net, substrate, params)
img = np.zeros((500, 500, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 500, 500), net, substrate=True)
cv2.imshow("NN", img)
cv2.waitKey(1)
if max(fitnesses) > 15.0:
break
# Epoch
generations = generation
pop.Epoch()
return generations
def reformat_sim_info(self):
# Pull out the Best Performer (Leader) Genotype and build its Phenotype
net = mneat.NeuralNetwork()
self.sim.pop.Species[0].GetLeader().BuildPhenotype(net)
# Network characteristics for visualization
# - NODES
out_node = [
net.connections[connection].source_neuron_idx
for connection in range(len(net.connections))
]
in_node = [
net.connections[connection].target_neuron_idx
for connection in range(len(net.connections))
]
nodes = list(set(out_node + in_node))
types = [net.neurons[node].type for node in range(len(net.neurons))]
# - CONNECTIONS
weights = [
net.connections[connection].weight
for connection in range(len(net.connections))
]
edges = zip(out_node, in_node)
return nodes, types, edges, weights
def run_network(genome, env, episode_count=1):
discrete_output = isinstance(env.action_space,
gym.spaces.discrete.Discrete)
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
cumulated_reward = 0
episode_count = 1
for _ in xrange(episode_count):
ob = env.reset()
net.Flush()
for _ in xrange(env.spec.timestep_limit):
net.Input(np.reshape(ob, -1))
for _ in xrange(1):
net.Activate()
o = net.Output()
if discrete_output:
action = np.argmax(o)
else:
assert (len(o) % 2 == 0)
action = np.array([
np.random.normal(o[i], abs(o[i + 1]))
for i in xrange(0, len(o), 2)
])
ob, reward, done, _ = env.step(action)
cumulated_reward += reward
if done:
break
return cumulated_reward
def evaluate(genome):
# this creates a neural network (phenotype) from the genome
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
# let's input just one pattern to the net, activate it once and get the output
net.Input([1.0, 0.0, 1.0])
net.Activate()
output = net.Output()
# the output can be used as any other Python iterable. For the purposes of the tutorial,
# we will consider the fitness of the individual to be the neural network that outputs constantly
# 0.0 from the first output (the second output is ignored)
#fitness = 1.0 - output[0]
fitness = 0
if output[0] >= 0.5:
output[0] = 1. - output[0]
fitness = 0.5 - (0.5 - output[0])
return fitness, output[0]
def getbest():
g = NEAT.Genome(0, 3, 0, 1, False,
NEAT.ActivationFunction.UNSIGNED_SIGMOID,
NEAT.ActivationFunction.UNSIGNED_SIGMOID, 0, params)
pop = NEAT.Population(g, params, True, 1.0)
generations = 0
for generation in range(1000):
genome_list = NEAT.GetGenomeList(pop)
fitness_list = NEAT.EvaluateGenomeList_Serial(genome_list,
evaluate,
display=False)
NEAT.ZipFitness(genome_list, fitness_list)
best = max([x.GetLeader().GetFitness() for x in pop.Species])
# print 'Best fitness:', best, 'Species:', len(pop.Species)
# test
net = NEAT.NeuralNetwork()
pop.Species[0].GetLeader().BuildPhenotype(net)
img = np.zeros((250, 250, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 250, 250), net)
cv2.imshow("nn_win", img)
cv2.waitKey(1)
pop.Epoch()
# print "Generation:", generation
generations = generation
if best > 15.5:
break
return generations
def evaluate(self, genotype, generation):
# this creates a neural network (phenotype) from the genome
net = NEAT.NeuralNetwork()
genotype.BuildPhenotype(net)
nn = neat_net_wrapper.NeatNetWrapper(net)
fitness_sum = 0.0
punishment = 3.0 # punishment for partial captures, small misses and large captures
for i in range(self.args.num_scenarios):
nn.flush()
seed = i + (997 * generation if self.args.mode == 'dynamic' else 0)
beer_tracker = BeerTracker(
nn=nn,
seed=seed
)
beer_tracker.run()
fitness_value = (
1 * beer_tracker.world.agent.num_small_captures +
(-punishment) * beer_tracker.world.agent.num_partial_captures +
(-punishment) * beer_tracker.world.agent.num_small_misses +
(-punishment) * beer_tracker.world.agent.num_large_captures
)
fitness_sum += fitness_value
fitness_sum /= self.args.num_scenarios
return fitness_sum
def testNetwork(data, n_classes, genome, evals, file, seed):
genome = pickle.loads(genome)
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
predictions = []
for example in data["X_test"]:
net.Flush()
net.Input(example)
for _ in range(3):
net.Activate()
result = softmax(net.Output())
guess = np.argmax(result)
predictions.append(guess)
cm = confusion_matrix(data["y_test"], predictions)
corr = 0
for i in xrange(n_classes):
for j in xrange(n_classes):
if i == j:
corr += cm[i, j]
acc = corr / float(len(predictions))
# dataset, architecture, seed, evals, acc
file.write("neat %d %d %f\n" % (seed, evals, acc))
def getbest(i):
g = NEAT.Genome(0, 3, 0, 1, False, NEAT.ActivationFunction.UNSIGNED_SIGMOID,
NEAT.ActivationFunction.UNSIGNED_SIGMOID, 0, params, 0)
pop = NEAT.Population(g, params, True, 1.0, i)
# pop.RNG.Seed(int(time.clock()*100))
pop.RNG.Seed(1234)
generations = 0
for generation in range(max_generations):
genome_list = NEAT.GetGenomeList(pop)
fitness_list = EvaluateGenomeList_Serial(genome_list, evaluate, display=False)
# fitness_list = EvaluateGenomeList_Parallel(genome_list, evaluate, display=False)
NEAT.ZipFitness(genome_list, fitness_list)
pop.Epoch()
generations = generation
best = max(fitness_list)
if best > 15.0:
break
net = NEAT.NeuralNetwork()
pop.GetBestGenome().BuildPhenotype(net)
# img = NEAT.viz.Draw(net)
# cv2.imshow("current best", img)
# cv2.waitKey(1)
return generations, net.NumHiddenNeurons(), net.NumConnections()
def validator(self, NEAT_file):
""" Validate a single run.
Args:
NEAT_File: file for the NEAT genome
"""
global man, quadruped
# Initialize the manager to be unique to the process.
man = ODEManager(near_callback,
stepsize=self.dt / self.n,
log_data=self.log_frames,
run_num=self.run_num)
# Initialize the quadruped
quadruped = Quadruped(man=man)
# If logging the output, tell manager to write the body type, dimensions, and position to the logging file.
if self.log_frames:
man.log_world_setup(self.eval_time, ind_num=self.run_num)
# Load in the best performing NEAT genome
genome = NEAT.Genome(NEAT_file)
self.current_network = NEAT.NeuralNetwork()
if not self.hyperNEAT:
genome.BuildPhenotype(self.current_network)
else:
genome.BuildHyperNEATPhenotype(self.current_network,
self.substrate)
fit = self.physics_only_simulation_validator()
print(fit)
def objF(genome):
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
random_state = np.random.get_state()
np.random.seed(generation)
sampled_data = np.random.choice(len(X_train),
num_samples,
replace=False)
np.random.set_state(random_state)
cur_data = X_train[sampled_data]
cur_label = y_train[sampled_data]
cum_correct = 0
for example, cor in zip(cur_data, cur_label):
net.Flush()
net.Input(example)
for _ in range(3):
net.Activate()
result = softmax(net.Output())
loss_sum = 0
for q, out in enumerate(result):
if q != cor:
loss_sum += max(0, out - result[int(cor)] + 1)
# Could also train with all results (e.g. l2 or logloss)
#if guess == cor:
#cum_correct += 1
cum_correct += loss_sum
# Return negative loss because fitness is expected
return -cum_correct
def evaluate_individ_obj_function(genome, generation):
"""
The function to evaluate individual objective function
Arguments:
genome: The objective function genome
generation: The current generation of evolution
Returns:
The NoveltyItem created using evaluation results.
"""
# create NoveltyItem for genome and store it into map
genome_id = genome.GetID()
n_item = archive.NoveltyItem(generation=generation, genomeId=genome_id)
# run the simulation
multi_net = NEAT.NeuralNetwork()
genome.BuildPhenotype(multi_net)
depth = 2
try:
genome.CalculateDepth()
depth = genome.GetDepth()
except:
pass
obj_net = ANN(multi_net, depth=depth)
# set inputs and get ouputs ([a, b])
output = obj_net.activate([0.5])
# store coefficients
n_item.data.append(output[0])
n_item.data.append(output[1])
return n_item
def evaluate_individual(self, genome):
""" Evaluate an individual solution.
Args:
genome: genome of the individual to evaluate
Returns:
fitness value of the individual
"""
global man, current_network, quadruped
# Initialize the manager to be unique to the process.
man = ODEManager(near_callback,
stepsize=self.dt / self.n,
log_data=self.log_frames)
# Initialize the quadruped
quadruped = Quadruped(man=man)
# Load in the ANN from the population
self.current_network = NEAT.NeuralNetwork()
if not self.hyperNEAT:
genome.BuildPhenotype(self.current_network)
else:
genome.BuildHyperNEATPhenotype(self.current_network,
self.substrate)
# Conduct the evaluation
fit = self.physics_only_simulation()
#print(genome.GetID(), fit)
return fit, len(self.current_network.neurons), len(
self.current_network.connections)
def simulate_genome(genome, traits, params, export_folder=None, verbose=False):
network = NEAT.NeuralNetwork()
genome.BuildPhenotype(network)
genome.CalculateDepth()
depth = genome.GetDepth()
return simulate_network(network, depth, traits, params, export_folder, verbose)
def evaluate(genome):
multi_net = NEAT.NeuralNetwork()
genome.BuildPhenotype(multi_net)
multi_net.Flush()
fitness = cart.eval_fitness(net=ANNWrapper(multi_net))
return fitness
def run_network(genome, cur_data, cur_label):
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
try:
actual_results = []
results = []
for example in cur_data:
net.Flush()
net.Input(example)
for _ in range(3):
net.Activate()
actual_result = net.Output()
result = softmax(actual_result)
# print "NEW"
# for element in actual_result:
# print element
# print result
actual_results.append([actual_result[0], actual_result[1]])
results.append(result)
predictions = [np.argmax(result) for result in results]
acc = accuracy_score(cur_label, predictions)
cost = log_loss(cur_label, results)
except:
print actual_results
print results
return acc, cost
def produce_output_sound(self, that_individual):
output_filename = '{0}.cross_adapted.{1}.wav'.format(
self.input_sound.filename, that_individual.get_id())
# this creates a neural network (phenotype) from the genome
net = NEAT.NeuralNetwork()
that_individual.genotype.BuildPhenotype(net)
output_vectors = []
for input_vector in self.neural_input_vectors:
net.Flush()
net.Input(input_vector)
net.Activate()
output = net.Output()
output = [min(1.0, max(0.0, x)) for x in output]
output_vectors.append(output)
that_individual.set_neural_output(zip(*output_vectors))
process, resulting_sound, csd_path = self.cross_adapt(
parameter_vectors=output_vectors,
effect=self.effect,
output_filename=output_filename)
return process, resulting_sound, csd_path
def getbest():
g = NEAT.Genome(0, substrate.GetMinCPPNInputs(), 0,
substrate.GetMinCPPNOutputs(), False,
NEAT.ActivationFunction.SIGNED_GAUSS,
NEAT.ActivationFunction.SIGNED_GAUSS, 0, params)
pop = NEAT.Population(g, params, True, 1.0)
for generation in range(1000):
genome_list = NEAT.GetGenomeList(pop)
# fitnesses = NEAT.EvaluateGenomeList_Parallel(genome_list, evaluate)
fitnesses = NEAT.EvaluateGenomeList_Serial(genome_list,
evaluate,
display=False)
[
genome.SetFitness(fitness)
for genome, fitness in zip(genome_list, fitnesses)
]
best = max([x.GetLeader().GetFitness() for x in pop.Species])
# print 'Best fitness:', best
# test
net = NEAT.NeuralNetwork()
pop.Species[0].GetLeader().BuildPhenotype(net)
img = np.zeros((250, 250, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 250, 250), net)
cv2.imshow("CPPN", img)
net = NEAT.NeuralNetwork()
pop.Species[0].GetLeader().BuildHyperNEATPhenotype(net, substrate)
img = np.zeros((250, 250, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 250, 250), net, substrate=True)
cv2.imshow("NN", img)
cv2.waitKey(1)
pop.Epoch()
# print "Generation:", generation
generations = generation
if best > 15.5:
break
return generations
def evaluate(genome):
multi_net = NEAT.NeuralNetwork()
genome.BuildPhenotype(multi_net)
multi_net.Flush()
fitness = cart.eval_fitness(net=ANNWrapper(multi_net),
action_evaluator=cart.two_ouputs_action_evaluator)# tanh_action_evaluator)
return fitness
def build_network_single(genome, method='neat', substrate=None, **kwargs):
net = neat.NeuralNetwork()
if method in ['neat', 'gdneat']:
genome.BuildPhenotype(net)
elif method == 'hyperneat':
genome.BuildHyperNEATPhenotype(net, substrate)
else:
raise ValueError('Invalid method: {}'.format(method))
return net
def getbest():
g = NEAT.Genome(0, 7, 1, False, NEAT.ActivationFunction.SIGNED_SIGMOID,
NEAT.ActivationFunction.SIGNED_SIGMOID, params)
pop = NEAT.Population(g, params, True, 1.0)
for generation in range(2000):
genome_list = NEAT.GetGenomeList(pop)
# fitnesses = NEAT.EvaluateGenomeList_Parallel(genome_list, evaluate)
fitnesses = NEAT.EvaluateGenomeList_Serial(genome_list,
evaluate_xor,
display=True)
[
genome.SetFitness(fitness)
for genome, fitness in zip(genome_list, fitnesses)
]
best = max([x.GetLeader().GetFitness() for x in pop.Species])
net = NEAT.NeuralNetwork()
pop.Species[0].GetLeader().BuildPhenotype(net)
img = np.zeros((500, 500, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 500, 500), net)
cv2.imshow("CPPN", img)
net = NEAT.NeuralNetwork()
pop.Species[0].GetLeader().Build_ES_Phenotype(net, substrate, params)
img = np.zeros((500, 500, 3), dtype=np.uint8)
img += 10
utilities.DrawPhenotype(img, (0, 0, 500, 500), net, substrate=True)
cv2.imshow("NN", img)
cv2.waitKey(1)
generations = generation
if best > 15.0:
break
pop.Epoch()
return generations
def evaluate_genome(env, genome, trials):
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
avg_reward = 0
for trial in range(trials):
f = do_trial(env, net, False)
avg_reward += f
avg_reward /= trials
fitness = 20 + avg_reward
return fitness
def Draw(x, size=(300, 300)):
img = np.zeros((size[0], size[1], 3), dtype=np.uint8)
img += 10
if isinstance(x, mneat.NeuralNetwork):
DrawPhenotype(img, (0, 0, 250, 250), x)
else:
nn = mneat.NeuralNetwork()
x.BuildPhenotype(nn)
DrawPhenotype(img, (0, 0, 250, 250), nn)
return img
def build_phenotype(self, current_genome):
"""
Constructs an agent phenotype from its genotype.
:param current_genome: agent genome.
:return: agent phenotype network.
"""
net = mneat.NeuralNetwork()
current_genome.BuildPhenotype(net)
return net
def evaluate(genome):
# create a neural network for the genome
net = NEAT.NeuralNetwork()
genome.BuildESHyperNEATPhenotype(net, substrate, params)
# create a board
board = Board(width, height)
canAddPiece = 1
current = 0
line = 0
maxPieces = 300
# while the game is not over
while (current < maxPieces):
# get a piece
pieceNum = random.randint(0, 6)
piece = Piece(pieceNum)
# pass board configuation & piece to the neural network
conf = board.getBoard().ravel()
c2, c3, c4, c5 = [piece.getPieceArray(i).ravel() for i in range(4)]
conf = np.concatenate([conf, c2, c3, c4, c5])
#conf = [tuple(row) for row in conf]
np.append(conf, pieceNum)
np.append(conf, 1.0) # bias
net.Input(conf)
net.Activate()
output = net.Output()
if PRINT:
print "output:", output[0], "outout", output[1]
## Found bug, output isnt always between 0 and 1
col = int(sum(output[0:10]))
rot = int(sum(output[10:14]))
print col, rot
##print "output1:", output[0], "output2:", output[1]
# update the board by the output we get
##print "rot: ", rot, "col: ", col
canAddPiece = board.addPiece(piece, rot, col)
line = line + board.clearRows()
if canAddPiece == -1:
#print board
break
current += 1
# evaluate fitness
#return line
print "Rows Cleared:", line, " Pieces: ", current
return current
def evaluate(genome):
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
net.Flush()
ge = GpuExec()
full_input = np.array([1, 0, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1], dtype=np.float32)
out = ge.eval(full_input, 3, 4, net)
# print(out)
targets = [1, 1, 0, 0]
err = np.abs(out - targets)
return (4 - np.sum(err)) ** 2
def get_signals(genome, df):
# this creates a neural network (phenotype) from the genome
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
signals = []
for index, row in df.iterrows():
row_list = list(row)
row_list.pop(0)
net.Input(row_list)
net.Activate()
signals.append(net.Output()[0])
return signals
def evaluate(genome):
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
avg_reward = 0
for trial in range(trials):
avg_reward += do_trial(net)
avg_reward /= trials
#print(avg_reward)
return 1000000 + avg_reward
def simGeneration(): genomeList = NEAT.GetGenomeList(pop) #make graph graph = mp.Map(20, 20) for g in genomeList: net = NEAT.NeuralNetwork() g.BuildPhenotype(net) fitness = evaluate(net, graph) g.SetFitness(fitness) pop.Epoch()
def evaluate_xor(genome):
net = NEAT.NeuralNetwork()
genome.BuildHyperNEATPhenotype(net, substrate)
# nn=genome.BuildHyperNEATPhenotype(net, substrate)
# error = 0
# depth = 5
# do stuff and return the fitness
# net.Flush()
net = NEAT.NeuralNetwork()
genome.BuildPhenotype(net)
img = np.zeros((400, 400, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 400, 400), net)
cv2.imshow("CPPN", img)
# Visualize best network's Pheotype
net = NEAT.NeuralNetwork()
genome.BuildESHyperNEATPhenotype(net, substrate, params)
img = np.zeros((800, 800, 3), dtype=np.uint8)
img += 10
NEAT.DrawPhenotype(img, (0, 0, 800, 800), net, substrate=True)
cv2.imshow("NN", img)
cv2.waitKey(33)
subprocess.call('./autostart.sh',shell=True)
# print("b")
# time.sleep(1)
# print("c")
main(MyDriver(net=net))
# print("d")
with open("mydata.txt",'r') as f:
fitt=f.read()
# os.system('pkill torcs')
subprocess.call('./autostop.sh',shell=True)
# print("fitness ******************* ",fitt)
return float(fitt)
