def generate(start):
u = Universe()
def fitness(f):
return u.do_simulation(f, False)
l = HillClimber(fitness, start, maxEvaluations=int(sys.argv[1]))
best, fitn = l.learn()
print(f"fitness: {fitn}")
#u.do_simulation(u.bestfaller, noStop=False)
u.record.finish()
from random import random
from pybrain.structure.evolvables.evolvable import Evolvable
from pybrain.optimization import HillClimber
class SimpleEvo(Evolvable):
def __init__(self, x):
self.x = max(0, min(x, 20))
def mutate(self):
self.x = max(0, min(self.x + random() - 0.3, 20))
def copy(self):
return SimpleEvo(self.x)
def randomize(self):
self.x = 20 * random()
def __repr__(self):
return '<-%.2f->' % (self.x)
if __name__ == "__main__":
x0 = SimpleEvo(1.2)
l = HillClimber(lambda x: x.x, x0, maxEvaluations=500)
result = l.learn()
print result
from pybrain.optimization import HillClimber, CMAES #@UnusedImport
# from pybrain.rl.learners.continuous.policygradients import ENAC
# from pybrain.rl.agents.learning import LearningAgent
from pybrain.rl.agents import OptimizationAgent
from pybrain.rl.experiments import EpisodicExperiment
# any episodic task
task = BalanceTask()
# any neural network controller
net = buildNetwork(task.outdim, 1, task.indim)
# any optimization algorithm to be plugged in, for example:
# learner = CMAES(storeAllEvaluations = True)
# or:
learner = HillClimber(storeAllEvaluations=True)
# in a non-optimization case the agent would be a LearningAgent:
# agent = LearningAgent(net, ENAC())
# here it is an OptimizationAgent:
agent = OptimizationAgent(net, learner)
# the agent and task are linked in an Experiment
# and everything else happens under the hood.
exp = EpisodicExperiment(task, agent)
exp.doEpisodes(100)
print('Episodes learned from:', len(learner._allEvaluations))
n, fit = learner._bestFound()
print('Best fitness found:', fit)
print('with this network:')
from pybrain.rl.environments.cartpole.balancetask import BalanceTask from pybrain.tools.shortcuts import buildNetwork from pybrain.rl.agents import OptimizationAgent from pybrain.rl.experiments import EpisodicExperiment from pybrain.optimization import HillClimber task = BalanceTask() net = buildNetwork(task.outdim, 3, task.indim) HillClimber(task, net, maxEvaluations=100).learn() agent = OptimizationAgent(net, HillClimber()) exp = EpisodicExperiment(task, agent) print(exp.doEpisodes(100))
def run_experiment():
# Create the controller network
HIDDEN_NODES = 4
RUNS = 2
BATCHES = 1
PRINTS = 1
EPISODES = 500
env = None
start_state_net = None
run_results = []
# Set up plotting tools for the experiments
tools = ExTools(BATCHES, PRINTS)
# Run the experiment
for run in range(RUNS):
if run == 0:
continue
# If an environment already exists, shut it down
if env:
env.closeSocket()
# Create the environment
env = create_environment()
# Create the task
task = Pa10MovementTask(env)
# Create the neural network. Only create the network once so it retains
# the same starting values for each run.
if start_state_net:
net = start_state_net.copy()
else:
# Create the initial neural network
net = create_network(
in_nodes=env.obsLen,
hidden_nodes=HIDDEN_NODES,
out_nodes=env.actLen
)
start_state_net = net.copy()
# Create the learning agent
learner = HillClimber(storeAllEvaluations=True)
agent = OptimizationAgent(net, learner)
tools.agent = agent
# Create the experiment
experiment = EpisodicExperiment(task, agent)
# Perform all episodes in the run
for episode in range(EPISODES):
experiment.doEpisodes(BATCHES)
# Calculate results
all_results = agent.learner._allEvaluations
max_result = np.max(all_results)
min_result = np.min(all_results)
avg_result = np.sum(all_results) / len(all_results)
run_results.append((run, max_result, min_result, avg_result))
# Make the results directory if it does not exist
if not os.path.exists(G_RESULTS_DIR):
os.mkdir(G_RESULTS_DIR)
# Write all results to the results file
with open(os.path.join(G_RESULTS_DIR, 'run_%d.txt' % run), 'w+') as f:
# Store the calculated max, min, avg
f.write('RUN, MAX, MIN, AVG\n')
f.write('%d, %f, %f, %f\n' % (run, max_result, min_result, avg_result))
# Store all results from this run
f.write('EPISODE, REWARD\n')
for episode, result in enumerate(all_results):
f.write('%d, %f\n' % (episode, result))
return