def get_re_experiment(case, minor=1):
""" Returns an experiment that uses the Roth-Erev learning method.
"""
gen = case.generators
profile = array([1.0])
maxSteps = len(profile)
experimentation = 0.55
recency = 0.3
tau = 100.0
decay = 0.99#9995
market = pyreto.SmartMarket(case, priceCap=cap, decommit=decommit,
auctionType=auctionType)
experiment = pyreto.continuous.MarketExperiment([], [], market, profile)
for g in gen[0:2]:
#learner = RothErev(experimentation, recency)
learner = VariantRothErev(experimentation, recency)
learner.explorer = BoltzmannExplorer(tau, decay)
task, agent = get_discrete_task_agent([g], market, nStates, nOffer,
markups, withholds, maxSteps, learner)
experiment.tasks.append(task)
experiment.agents.append(agent)
task1, agent1 = get_zero_task_agent(gen[2:3], market, nOffer, maxSteps)
experiment.tasks.append(task1)
experiment.agents.append(agent1)
return experiment
def get_re_experiment(case, minor=1):
""" Returns an experiment that uses the Roth-Erev learning method.
"""
locAdj = "ac"
experimentation = 0.55
recency = 0.3
tau = 100.0
decay = 0.999
nStates = 3 # stateless RE?
Pd0 = get_pd_max(case, profile)
Pd_min = get_pd_min(case, profile)
market = pyreto.SmartMarket(case, priceCap=cap, decommit=decommit,
auctionType=auctionType,
locationalAdjustment=locAdj)
experiment = pyreto.continuous.MarketExperiment([], [], market)
portfolios, sync_cond = get_portfolios3()
for gidx in portfolios:
g = [case.generators[i] for i in gidx]
learner = VariantRothErev(experimentation, recency)
learner.explorer = BoltzmannExplorer(tau, decay)
task, agent = get_discrete_task_agent(g, market, nStates, nOffer,
markups, withholds, maxSteps, learner, Pd0, Pd_min)
print "ALL ACTIONS:", len(task.env._allActions) * nStates
experiment.tasks.append(task)
experiment.agents.append(agent)
passive = [case.generators[i] for i in sync_cond]
passive[0].p_min = 0.001 # Avoid invalid offer withholding.
passive[0].p_max = 0.002
task, agent = get_zero_task_agent(passive, market, 1, maxSteps)
experiment.tasks.append(task)
experiment.agents.append(agent)
return experiment
def case6ww1():
case = pylon.Case.load(join("..", "data", "case6ww.pkl"))
case.generators[0].p_cost = (0.0, 5.0, 200.0)
case.generators[1].p_cost = (0.0, 6.5, 200.0)
case.generators[2].p_cost = (0.0, 2.0, 200.0)
case.generators[0].c_shutdown = 100.0
#case.generators[0].p_min = 0.0 # TODO: Unit-decommitment.
#case.generators[1].p_min = 0.0
##case.generators[2].p_min = 0.0
case.generators[0].p_max = 100.0
case.generators[1].p_max = 70.0
case.generators[2].p_max = 70.0
vre = VariantRothErev(experimentation=0.55, recency=0.3)
vre.explorer = BoltzmannExplorer()#tau=100, decay=0.95)
learners = [vre, Q(), Reinforce()]
profile = [0.9, 0.6]
m = (20, 75) # markups
nb = 1 # no. offers
ns = 3 # no. states
mx = 60.0 # max markup
weeks = 2
days = 2
outdir = "/tmp/case6ww1"
dc = True
trials = 1
for i in range(trials):
roundrobin(case, learners, profile, m, nb, ns, mx, weeks, days,
outdir, dc, trial=i)
Pd_min = get_pd_min(case, profile)
market = pyreto.SmartMarket(case,
priceCap=cap,
decommit=decommit,
auctionType=auctionType,
locationalAdjustment=locAdj)
experiment = pyreto.continuous.MarketExperiment([], [], market)
portfolios, sync_cond = get_portfolios3()
for gidx in portfolios:
g = [case.generators[i] for i in gidx]
learner = VariantRothErev(experimentation, recency)
learner.explorer = BoltzmannExplorer(tau, decay)
task, agent = get_discrete_task_agent(g, market, nStates, nOffer,
markups, withholds, maxSteps,
learner, Pd0, Pd_min)
print "ALL ACTIONS:", len(task.env._allActions) * nStates
experiment.tasks.append(task)
experiment.agents.append(agent)
passive = [case.generators[i] for i in sync_cond]
passive[0].p_min = 0.001 # Avoid invalid offer withholding.
passive[0].p_max = 0.002
task, agent = get_zero_task_agent(passive, market, 1, maxSteps)
distrib = scipy.array([[0.7, 0.2, 0.1],
[0.1, 0.6, 0.3],
[0.4, 0.2, 0.3],
[0.5, 0.45, 0.05],
[0.3, 0.5, 0.2]])
env = BanditEnvironment(payouts, distrib)
task = BanditTask(env)
table = PropensityTable(payouts.shape[0])
table.initialize(500.0)
#learner = RothErev(experimentation=0.55, recency=0.3)
learner = VariantRothErev(experimentation=0.65, recency=0.3)
learner.explorer = BoltzmannExplorer(tau=100.0, decay=0.9995)
agent = LearningAgent(table, learner)
experiment = Experiment(task, agent)
epis = int(1e1)
batch = 2
avgRewards = scipy.zeros(epis)
allActions = scipy.zeros(epis * batch)
c = 0
for i in range(epis):
experiment.doInteractions(batch)
avgRewards[i] = scipy.mean(agent.history["reward"])
allActions[c:c + batch] = agent.history["action"].flatten() + 1
case = pylon.Case.load(join("..", "data", "case6ww.pkl"))
case.generators[0].p_cost = (0.0, 5.0, 200.0)
case.generators[1].p_cost = (0.0, 6.5, 200.0)
case.generators[2].p_cost = (0.0, 2.0, 200.0)
case.generators[0].c_shutdown = 100.0
#case.generators[0].p_min = 0.0 # TODO: Unit-decommitment.
#case.generators[1].p_min = 0.0
##case.generators[2].p_min = 0.0
case.generators[0].p_max = 100.0
case.generators[1].p_max = 70.0
case.generators[2].p_max = 70.0
vre = VariantRothErev(experimentation=0.55, recency=0.3)
vre.explorer = BoltzmannExplorer() #tau=100, decay=0.95)
learners = [vre, Q(), Reinforce()]
profile = [0.9, 0.6]
m = (20, 75) # markups
nb = 1 # no. offers
ns = 3 # no. states
mx = 60.0 # max markup
weeks = 2
days = 2
outdir = "/tmp/case6ww1"
[150.0, 50.0, 1000.0], # Expected value: 147.5
[700.0, 800.0, 900.0]
]) # Expected value: 790
distrib = scipy.array([[0.7, 0.2, 0.1], [0.1, 0.6, 0.3], [0.4, 0.2, 0.3],
[0.5, 0.45, 0.05], [0.3, 0.5, 0.2]])
env = BanditEnvironment(payouts, distrib)
task = BanditTask(env)
table = PropensityTable(payouts.shape[0])
table.initialize(500.0)
#learner = RothErev(experimentation=0.55, recency=0.3)
learner = VariantRothErev(experimentation=0.65, recency=0.3)
learner.explorer = BoltzmannExplorer(tau=100.0, decay=0.9995)
agent = LearningAgent(table, learner)
experiment = Experiment(task, agent)
epis = int(1e1)
batch = 2
avgRewards = scipy.zeros(epis)
allActions = scipy.zeros(epis * batch)
c = 0
for i in range(epis):
experiment.doInteractions(batch)
avgRewards[i] = scipy.mean(agent.history["reward"])
allActions[c:c + batch] = agent.history["action"].flatten() + 1
