# "ENV_NAME": "CartPole-v0",
params = {
"PLATFORM": "openai",
"ENV_NAME": "MountainCarContinuous-v0",
"METHOD": "QLearning",
"REPORTING_INTERVAL": 100,
"LOG_LEVEL": 2,
"NUMBER_EPISODES_MEAN": 10,
"MEAN_REWARD_BOUND": 90,
"NUM_TRIALS": 1,
"MAX_EPISODES": 10000,
"EPSILON_DECAY_LAST_FRAME": 1000000, # 500000 do not solve
"EPSILON_START": 1.0,
"EPSILON_FINAL": 0.02, # 0.02
"LEARNING_RATE": 0.05,
"GAMMA": 0.99,
"DISCRETIZE_STATE": True,
"DISCRETIZE_STATE_BIN_SIZE": 10,
"DISCRETIZE_ACTION": True,
"DISCRETIZE_ACTION_BIN_SIZE": 50,
}
exp = UntilWinExperiment(params)
exp.run()
# solved the problem using action discretization
# Problem solved in 1839 episodes
# Trial took 62.19 seconds
dqn_focus_sharing = dqn_sharing.copy()
dqn_focus_sharing.update(focus)
# others
dqn_prio_sharing = dqn_sharing.copy()
dqn_prio_sharing.update(prio)
dqn_prio_focus_sharing = dqn_sharing.copy()
dqn_prio_focus_sharing.update(prio)
dqn_prio_focus_sharing.update(focus)
## prepare the experiment
exp_group = exp_group
experiments = {
'dqn': dqn,
'dqn_prio': dqn_prio,
'dqn_sharing': dqn_sharing,
'dqn_prio_sharing': dqn_prio_sharing,
'dqn_focus_sharing': dqn_focus_sharing,
'dqn_prio_focus_sharing': dqn_prio_focus_sharing,
}
for exp_name, params in reversed(list(experiments.items())):
print(exp_name, params)
if 'sharing' in exp_name:
exp = MultiAgentExperiment(params, exp_name, exp_group)
else:
exp = UntilWinExperiment(params, exp_name, exp_group)
exp.run()
"EPSILON_START": 1.0,
"EPSILON_FINAL": 0,
"LEARNING_RATE": 0.3,
"GAMMA": 0.99
}
params["ENV_NAME"] = "FrozenLakeNotSlippery-v0"
results = []
methods = [
"QLearning", "Sarsa", "FirstVisitMonteCarlo", "EveryVisitMonteCarlo",
"NStepsQLearning", "NStepsSarsa"
]
for method in methods:
params["METHOD"] = method
exp = UntilWinExperiment(params)
result = exp.run()
results.append(result)
for method, result in zip(methods, results):
print("Method {} took an average of {:.2f} episodes".format(
method, result))
# Results, using all same parameters:
# For "FrozenLakeNotSlippery-v0"
#
# Method QLearning took an average of 321.20 episodes
# Method Sarsa took an average of 500.53 episodes
# Method FirstVisitMonteCarlo took an average of 341.27 episodes
# Method EveryVisitMonteCarlo took an average of 304.73 episodes
# Method NStepsQLearning took an average of 287.47 episodes