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)
def pIteration(world, userMap, maxX, maxY, discount=0.99, MAX_ITERATIONS=100): 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) policy_converged = defaultdict(list) last_policy = defaultdict(list) allStates = getAllStates(domain, rf, tf, initialState) print("*** {} Policy Iteration Analysis".format(world)) MAX_ITERATIONS = MAX_ITERATIONS iterations = range(1, MAX_ITERATIONS + 1) pi = PolicyIteration(domain,rf,tf,discount,hashingFactory,-1,1,1); pi.setDebugCode(0) for nIter in iterations: startTime = clock() #pi = PolicyIteration(domain,rf,tf,discount,hashingFactory,-1,1, nIter); #pi.setDebugCode(0) # run planning from our initial state p = pi.planFromState(initialState); endTime = clock() timing['Policy'].append((endTime-startTime)*1000) convergence['Policy'].append(pi.lastPIDelta) # evaluate the policy with one roll out visualize the trajectory runEvals(initialState, p, rewards['Policy'], steps['Policy'], rf, tf, evalTrials=1) if nIter == 1 or nIter == 50: simpleValueFunctionVis(pi, p, initialState, domain, hashingFactory, "Policy Iteration{}".format(nIter)) policy = pi.getComputedPolicy() allStates = pi.getAllStates() current_policy = [[(action.ga, action.pSelection) for action in policy.getActionDistributionForState(state)] for state in allStates] policy_converged['Policy'].append(current_policy == last_policy) last_policy = current_policy simpleValueFunctionVis(pi, p, initialState, domain, hashingFactory, "Policy Iteration{}".format(nIter)) dumpPolicyMap(MapPrinter.printPolicyMap(allStates, p, gen.getMap()), world + ' Policy Iteration Policy Map.pkl') dumpCSVp(iterations, timing['Policy'], rewards['Policy'], steps['Policy'],convergence['Policy'], world, 'Policy', policy_converged['Policy'])
def vIteration(world, userMap, maxX, maxY, discount=0.99, MAX_ITERATIONS=100): 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) print("*** {} Value Iteration Analysis".format(world)) MAX_ITERATIONS = MAX_ITERATIONS iterations = range(1, MAX_ITERATIONS + 1) vi = ValueIteration(domain, rf, tf, discount, hashingFactory, -1, 1); vi.setDebugCode(0) vi.performReachabilityFrom(initialState) vi.toggleUseCachedTransitionDynamics(False) timing['Value'].append(0) for nIter in iterations: startTime = clock() vi.runVI() p = vi.planFromState(initialState); endTime = clock() timing['Value'].append((endTime-startTime)*1000) convergence['Value'].append(vi.latestDelta) # evaluate the policy with evalTrials roll outs runEvals(initialState, p, rewards['Value'], steps['Value'], rf, tf, evalTrials=1) if nIter == 1 or nIter == 50: simpleValueFunctionVis(vi, p, initialState, domain, hashingFactory, "Value Iteration {}".format(nIter)) simpleValueFunctionVis(vi, p, initialState, domain, hashingFactory, "Value Iteration {}".format(nIter)) dumpPolicyMap(MapPrinter.printPolicyMap(allStates, p, gen.getMap()), world + ' Value Iteration Policy Map.pkl') dumpCSV(nIter, timing['Value'][1:], rewards['Value'], steps['Value'], convergence['Value'], world, 'Value')
[0, 1, 0, 0, 1, 1, 1, 1, 0, 0], [0, 1, 0, 0, 0, 0, 0, 1, 0, 0], [0, 1, 0, 0, 1, 0, 0, 1, 0, 0], [0, 1, 0, 0, 1, 0, 0, 1, 0, 0], [0, 1, 0, 0, 1, 0, 0, 1, 0, 0], [0, 1, 0, 0, 1, 0, 0, 1, 0, 0], [0, 1, 0, 0, 1, 0, 0, 1, 0, 0], [0, 1, 0, 1, 1, 0, 0, 1, 0, 0], [0, 0, 0, -3, 0, 0, 0, 0, 0, 0]] n = len(userMap) tmp = java.lang.reflect.Array.newInstance(java.lang.Integer.TYPE, [n, n]) for i in range(n): for j in range(n): tmp[i][j] = userMap[i][j] userMap = MapPrinter().mapToMatrix(tmp) maxX = maxY = n - 1 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) # Print the map that is being analyzed print "/////{} Grid World Analysis/////\n".format(world) MapPrinter().printMap(MapPrinter.matrixToMap(userMap)) visualizeInitialGridWorld(domain, gen, env) hashingFactory = SimpleHashableStateFactory() increment = MAX_ITERATIONS / NUM_INTERVALS timing = defaultdict(list) rewards = defaultdict(list) steps = defaultdict(list)