def qLearning(world, userMap, maxX, maxY, discount=0.9, MAX_ITERATIONS=1000):
gen = BasicGridWorld(userMap, maxX, maxY)
domain = gen.generateDomain()
initialState = gen.getExampleState(domain);
rf = BasicRewardFunction(maxX, maxY, userMap)
tf = BasicTerminalFunction(maxX, maxY)
env = SimulatedEnvironment(domain, rf, tf, initialState)
visualizeInitialGridWorld(domain, gen, env)
hashingFactory = SimpleHashableStateFactory()
timing = defaultdict(list)
rewards = defaultdict(list)
steps = defaultdict(list)
convergence = defaultdict(list)
allStates = getAllStates(domain, rf, tf, initialState)
MAX_ITERATIONS = MAX_ITERATIONS
NUM_INTERVALS = MAX_ITERATIONS;
iterations = range(1, MAX_ITERATIONS + 1)
qInit = 0
for lr in [0.01, 0.1, 0.5]:
for epsilon in [0.3, 0.5, 0.7]:
last10Chg = deque([10] * 10, maxlen=10)
Qname = 'Q-Learning L{:0.2f} E{:0.1f}'.format(lr, epsilon)
#agent = QLearning(domain, discount, hashingFactory, qInit, lr, epsilon, 300)
agent = QLearning(domain, discount, hashingFactory, qInit, lr, epsilon)
agent.setDebugCode(0)
print("*** {}: {}".format(world, Qname))
for nIter in iterations:
if nIter % 200 == 0:
print('Iteration: {}'.format(nIter))
startTime = clock()
#ea = agent.runLearningEpisode(env, 300)
ea = agent.runLearningEpisode(env)
env.resetEnvironment()
agent.initializeForPlanning(rf, tf, 1)
p = agent.planFromState(initialState) # run planning from our initial state
endTime = clock()
timing[Qname].append((endTime-startTime)*1000)
last10Chg.append(agent.maxQChangeInLastEpisode)
convergence[Qname].append(sum(last10Chg)/10.)
# evaluate the policy with one roll out visualize the trajectory
runEvals(initialState, p, rewards[Qname], steps[Qname], rf, tf, evalTrials=1)
if nIter % 1000 == 0:
dumpPolicyMap(MapPrinter.printPolicyMap(allStates, p, gen.getMap()),
'{} {} Iter {} Policy Map.pkl'.format(world, Qname, nIter))
simpleValueFunctionVis(agent, p, initialState, domain, hashingFactory, Qname)
dumpCSV(nIter, timing[Qname], rewards[Qname], steps[Qname], convergence[Qname], world, Qname)
for nIter in iterations:
if nIter % 50 == 0: print(nIter)
startTime = clock()
ea = agent.runLearningEpisode(env, 300)
# if len(timing[Qname])> 0:
# timing[Qname].append(timing[Qname][-1]+clock()-startTime)
# else:
#timing[Qname].append((clock()-startTime) * 1000)
if len(timing[Qname]) > 0:
timing[Qname].append(timing[Qname][-1] + clock() -
startTime)
else:
timing[Qname].append(clock() - startTime)
env.resetEnvironment()
agent.initializeForPlanning(rf, tf, 1)
p = agent.planFromState(
initialState) # run planning from our initial state
last10Chg.append(agent.maxQChangeInLastEpisode)
convergence[Qname].append(sum(last10Chg) / 10.)
# evaluate the policy with one roll out visualize the trajectory
runEvals(initialState, p, rewards[Qname], steps[Qname])
# if nIter == 50 :
# dumpPolicyMap(MapPrinter.printPolicyMap(allStates, p, gen.getMap()),'QL {} {} Iter {} Policy Map.pkl'.format(Qname,world,nIter))
if convergence[Qname][-1] < 0.5:
#dumpPolicyMap(MapPrinter.printPolicyMap(allStates, p, gen.getMap()),'QL {} {} Iter {} Policy Map.pkl'.format(Qname,world,nIter))
if flag:
simpleValueFunctionVis(agent, p, initialState,
domain, hashingFactory,
Qname + ' {}'.format(nIter))
flag = False