def plot_results_from_dump(envName, tmax, tplot):
"""
Requires result files to be named "results/cumRegret_" + envName + "_" + learner.name() + "_" + str(tmax)"
:param envName: name of the environment
:param tmax: assumes the data have been generated with maximal time horizon tmax.
:param tplot: the results are plotted only until time horizon tplot.
:return:
"""
if (envName in bW.registerWorlds.keys()):
regName = (bW.registerWorlds[envName])(0)
#print("Environment " + envName + " registered as " + regName)
envName = regName
envOpt = bW.makeWorld(envName)
median = []
quantile1 = []
quantile2 = []
learners = []
names = []
skip = max(1, (tmax // 1000))
itimes = [t for t in range(0, tmax, skip)]
times = [itimes[i] for i in range(len(itimes)) if i * skip < tplot]
#Declare list of algorithms (only to get their names):
learners.append(
ucrl.UCRL2(envOpt.observation_space.n,
envOpt.action_space.n,
delta=0.05))
learners.append(
c_ucrl.C_UCRL2(envOpt.observation_space.n,
envOpt.action_space.n,
envOpt,
envOpt.env.classes,
delta=0.05))
for learner in learners:
names.append(learner.name())
filename = "results/cumRegret_" + envName + "_" + learner.name(
) + "_" + str(tmax)
file = open(filename, 'rb')
data_j = pickle.load(file)
file.close()
q = np.quantile(data_j, 0.5, axis=0)
median.append([q[i] for i in range(len(q)) if i * skip < tplot])
q = np.quantile(data_j, 0.25, axis=0)
quantile1.append([q[i] for i in range(len(q)) if i * skip < tplot])
q = np.quantile(data_j, 0.75, axis=0)
quantile2.append([q[i] for i in range(len(q)) if i * skip < tplot])
plotCumulativeRegretsFromDump(names, envName, median, quantile1, quantile2,
times, tplot)
def demo_animate():
testName = 'riversail_10'
envName = (bW.registerWorlds[testName])(0)
env = bW.makeWorld(envName)
# -> Choose which learner to use:
# learner = lr.Random(env)
# learner = lh.Human(env)
learner = ucrl.UCRL2(env.observation_space.n,
env.action_space.n,
delta=0.05)
# learner = c_ucrl.C_UCRL2(env.observation_space.n, env.action_space.n, env, env.env.classes, delta=0.05)
animate(env, learner, 50, 'maze')
def multicoreXpsNoRender(envName, learner, nbReplicates, timeHorizon):
num_cores = multiprocessing.cpu_count()
envs = []
learners = []
timeHorizons = []
for i in range(nbReplicates):
envs.append(bW.makeWorld(envName))
learners.append(copy.deepcopy(learner))
timeHorizons.append(copy.deepcopy(timeHorizon))
t0 = time.time()
cumRewards = Parallel(n_jobs=num_cores)(
delayed(starOneXp)(i) for i in zip(envs, learners, timeHorizons))
elapsed = time.time() - t0
return cumRewards, elapsed / nbReplicates
def run_large_exp(envName="riverSwim", timeHorizon=1000, nbReplicates=100):
if (envName in bW.registerWorlds.keys()):
regName = (bW.registerWorlds[envName])(0)
print("Environment " + envName + " registered as " + regName)
envName = regName
envOpt = bW.makeWorld(envName)
print("Computing an estimate of the optimal policy (for regret)...")
opti_learner = opt.Opti_controller(envOpt.env, envOpt.observation_space.n,
envOpt.action_space.n)
print(opti_learner.policy)
print("*********************************************")
dump_cumRewardsAlgos = []
names = []
meanelapsedtimes = []
learners = []
learners.append(
ucrl.UCRL2(envOpt.observation_space.n,
envOpt.action_space.n,
delta=0.05))
learners.append(
c_ucrl.C_UCRL2(envOpt.observation_space.n,
envOpt.action_space.n,
envOpt,
envOpt.env.classes,
delta=0.05))
for learner in learners:
names.append(learner.name())
dump_cumRewards, meanelapsedtime = multicoreXpsNoRenderWithDump(
envName, learner, nbReplicates, timeHorizon)
dump_cumRewardsAlgos.append(dump_cumRewards)
meanelapsedtimes.append(meanelapsedtime)
opttimeHorizon = min(max((10000, timeHorizon)), 10**8)
cumReward_opti = oneXpNoRender(envOpt, envName, opti_learner,
opttimeHorizon)
gain = cumReward_opti[-1] / len(cumReward_opti)
print("Average gain is ", gain)
opti_reward = [[t * gain for t in range(timeHorizon)]]
filename = "results/cumMeans_" + envName + "_" + opti_learner.name(
) + "_" + str(timeHorizon) + "_" + str(time.time())
file = open(filename, 'wb')
pickle.dump(opti_reward, file)
file.close()
dump_cumRewardsAlgos.append(filename)
[
print(str(names[i]), "average runtime is ", meanelapsedtimes[i])
for i in range(len(names))
]
median, quantile1, quantile2, times = analyzeResults(
names, dump_cumRewardsAlgos, timeHorizon, envName)
plotCumulativeRegretsFromDump(names, envName, median, quantile1, quantile2,
times, timeHorizon)
print("*********************************************")