def coevo():
# Create a pool for the policies
pi_pool = Pool.spawn(Genome.open(NN_STRUCTURE_FILE), 20, std=8)
# Create a pool of z's, starting around [0.5,0.5], should probably be better
z_list = [[x] for x in np.linspace(0, 0.5, 5)]
genomes = BasicGenome.from_list(z_list, 5)
org_list = [Organism(genome) for genome in genomes]
z_pool = Pool(org_list)
avg_fitness = []
champ_fitness = []
for i in xrange(150):
pi_pool = eonn.epoch(pi_pool, len(pi_pool))
z_pool = eonn.epoch(z_pool, len(z_pool))
for pi_org, z_org in itertools.product(pi_pool, z_pool):
reward = cliff(pi_org.genome, z=[z_org.weights[0]], verbose=False)
pi_org.evals.append(reward)
z_org.evals.append(reward)
for org in z_pool:
org.evals = [np.var(org.evals)]
avg_fitness.append(pi_pool.fitness)
champion = max(pi_pool)
champ_fitness.append(champion.fitness)
return avg_fitness, champ_fitness
def do_evolution(pi_pool, z_pool, GP):
"""
Evolve the organisms and predict their values according to the GP
"""
# Evolve pools
pi_pool = eonn.epoch(pi_pool, len(pi_pool))
z_pool = eonn.epoch(z_pool, len(z_pool))
# Create prediction matrix for GP
x_predict = [
np.append(pi_org.weights, z_org.weights) for pi_org in pi_pool
for z_org in z_pool
]
# Get rewards and MSE
reward_predict, MSE = GP.predict(x_predict, eval_MSE=True)
return pi_pool, z_pool, x_predict, reward_predict, MSE
def find_best(GP, epochs=100):
""" Find the best policy in the GP """
pool = Pool.spawn(Genome.open(NN_STRUCTURE_FILE), 50, std=8)
all_z = list(np.linspace(0, max_wind, 10))
for n in xrange(epochs):
if n != 0:
pool = eonn.epoch(pool, len(pool))
weights = [np.append(org.weights, z) for org in pool for z in all_z]
reward = GP.predict(weights)
for i in xrange(len(pool)):
pool[i].evals = list(reward[i * len(all_z):(i + 1) * len(all_z)])
champion = max(pool)
return champion
def find_best_upper(GP, epochs=100):
""" Find policy with highest upperbound """
pool = Pool.spawn(Genome.open(NN_STRUCTURE_FILE), 50, std=8)
all_z = list(np.linspace(0, max_wind, 10))
for n in xrange(epochs):
if n != 0:
pool = eonn.epoch(pool, len(pool))
weights = [np.append(org.weights, z) for org in pool for z in all_z]
reward, MSE = GP.predict(weights, eval_MSE=True)
reward += 1.96 * np.sqrt(MSE)
for i in xrange(len(pool)):
pool[i].evals = list(reward[i * len(all_z):(i + 1) * len(all_z)])
champion = max(pool)
return champion