from training import LinearFATraining
from learners import SARSALambda_LinFA_ReplacingTraces
task = LinearFATileCoding3456BalanceTaskRewardBipolar()
learner = SARSALambda_LinFA_ReplacingTraces(task.nactions, task.outdim)
learner._lambda = 0.95
task.discount = learner.rewardDiscount
agent = LinearFA_Agent(learner)
agent.epsilonGreedy = True
agent.init_exploration = 0.5
# The state has a huge number of dimensions, and the logging causes me to run
# out of memory. We needn't log, since learning is done online.
agent.logging = False
performance_agent = LinearFA_Agent(learner)
performance_agent.logging = False
performance_agent.greedy = True
performance_agent.learning = False
experiment = EpisodicExperiment(task, agent)
# TODO PyBrain says that the learning rate needs to decay, but I don't see that
# described in Randlov's paper.
# A higher number here means the learning rate decays slower.
learner.learningRateDecay = 2000
# NOTE increasing this number above from the default of 100 is what got the
# learning to actually happen, and fixed the bug/issue where the performance
# agent's performance stopped improving.
tr = LinearFATraining(
'balance_sarsalambda_linfa_replacetrace_anneal_RewardBipolar_take2',
experiment,
performance_agent,
from training import LinearFATraining
from learners import SARSALambda_LinFA_ReplacingTraces
task = LinearFATileCoding3456BalanceTaskRewardPower8()
learner = SARSALambda_LinFA_ReplacingTraces(task.nactions, task.outdim)
learner._lambda = 0.95
task.discount = learner.rewardDiscount
agent = LinearFA_Agent(learner)
agent.epsilonGreedy = True
agent.init_exploration = 0.5
# The state has a huge number of dimensions, and the logging causes me to run
# out of memory. We needn't log, since learning is done online.
agent.logging = False
performance_agent = LinearFA_Agent(learner)
performance_agent.logging = False
performance_agent.greedy = True
performance_agent.learning = False
experiment = EpisodicExperiment(task, agent)
# TODO PyBrain says that the learning rate needs to decay, but I don't see that
# described in Randlov's paper.
# A higher number here means the learning rate decays slower.
learner.learningRateDecay = 2000
# NOTE increasing this number above from the default of 100 is what got the
# learning to actually happen, and fixed the bug/issue where the performance
# agent's performance stopped improving.
tr = LinearFATraining('balance_sarsalambda_linfa_replacetrace_anneal_RewardPower8_take1', experiment,
performance_agent, verbose=True)
tr.train(7000, performance_interval=1, n_performance_episodes=1,
delta_dist = dist_to_goal - dist_to_goal_last
return -delta_tilt - delta_dist * 0.01
task = LSPI_task()
learner = LSPI(9, 20)
task.rewardDiscount = 0.8
learner.rewardDiscount = 0.8
agent = LinearFA_Agent(learner)
agent.epsilonGreedy = True
exp = EpisodicExperiment(task, agent)
learner.learningRateDecay = 3000
max_agent = LinearFA_Agent(learner)
max_agent.learnerning = False
max_agent.greedy = True
task.env.saveWheelContactTrajectories(True)
plt.ion()
plt.figure(figsize=(8, 4))
ax1 = plt.subplot(1, 2, 1)
ax2 = plt.subplot(1, 2, 2)
def update_wheel_trajectories():
front_lines = ax2.plot(task.env.get_xfhist(), task.env.get_yfhist(), 'r')
back_lines = ax2.plot(task.env.get_xbhist(), task.env.get_ybhist(), 'b')
plt.axis('equal')