def createBanditInstancesAndSimulate(params,n_mc_sim):
n_sim = n_mc_sim
for i in range(len(params)):
results = []
time_horizon = params[i]['time_horizon']
number_of_arms = params[i]['number_of_arms']
number_of_exploration_per_arm = params[i]['number_of_exploration_per_arm']
exp_agent = ExploreThenExploit(time_horizon,number_of_arms,number_of_exploration_per_arm)
epsilon_greedy_constant_half_epsilonAgent = EpsilonGreedy(time_horizon,number_of_arms,[1/2]*time_horizon)
epsilon_greedy_constant_epsilonAgent = EpsilonGreedy(time_horizon,number_of_arms,[number_of_exploration_per_arm*number_of_arms/time_horizon]*time_horizon)
ubc_agent = UBC1Agent(time_horizon,number_of_arms)
se_agent = SuccessiveEliminationAgent(time_horizon,number_of_arms)
random_agent = Agent()
bandit = Bandit(time_horizon,number_of_arms,random_agent)
results.append(mc_simulate(n_sim,bandit))
bandit = Bandit(time_horizon,number_of_arms,exp_agent)
results.append(mc_simulate(n_sim,bandit))
bandit = Bandit(time_horizon,number_of_arms,epsilon_greedy_constant_epsilonAgent)
results.append(mc_simulate(n_sim,bandit,"constant-epsilon=rate-of-explore-exploit"))
bandit = Bandit(time_horizon,number_of_arms,epsilon_greedy_constant_half_epsilonAgent)
results.append(mc_simulate(n_sim,bandit,"constant-epsilon=0.5"))
bandit = Bandit(time_horizon,number_of_arms,ubc_agent)
results.append(mc_simulate(n_sim,bandit))
bandit = Bandit(time_horizon,number_of_arms,se_agent)
results.append(mc_simulate(n_sim,bandit))
plot(results,time_horizon,params[i])
def experiment2():
params = [{
"time_horizon" : 500,
"number_of_arms" : 5
},
{
"time_horizon" : 5000,
"number_of_arms" : 5
},
{
"time_horizon" : 500,
"number_of_arms" : 10
},
{
"time_horizon" : 5000,
"number_of_arms" : 10
},
{
"time_horizon" : 500,
"number_of_arms" : 20
},
{
"time_horizon" : 5000,
"number_of_arms" : 20
},
]
for i in range(len(params)):
results = []
time_horizon = params[i]["time_horizon"]
number_of_arms = params[i]["number_of_arms"]
agent1 = agentFactory("random",time_horizon,number_of_arms)
epsilons = []
for j in range(time_horizon):
epsilons.append(math.pow((j+1)*number_of_arms*math.log(j+1),1/3))
agent2 = agentFactory("epsilon-greedy",time_horizon,number_of_arms,epsilons)
agent3 = agentFactory("explore-then-exploit",time_horizon,number_of_arms,time_horizon/100)
agent4 = agentFactory("ucb1",time_horizon,number_of_arms)
agent5 = agentFactory("successive-elimination",time_horizon,number_of_arms)
bandit = Bandit(time_horizon,number_of_arms,agent1)
results.append(mc_simulate(n_sim,bandit))
bandit = Bandit(time_horizon,number_of_arms,agent2)
results.append(mc_simulate(n_sim,bandit))
bandit = Bandit(time_horizon,number_of_arms,agent3)
results.append(mc_simulate(n_sim,bandit,"N=T/100"))
bandit = Bandit(time_horizon,number_of_arms,agent4)
results.append(mc_simulate(n_sim,bandit))
bandit = Bandit(time_horizon,number_of_arms,agent5)
results.append(mc_simulate(n_sim,bandit))
plot(results,time_horizon,params[i])
def experiment1():
params = [{
"time_horizon" : 1000,
"number_of_arms" : 5
},
{
"time_horizon" : 10000,
"number_of_arms" : 5
},
{
"time_horizon" : 1000,
"number_of_arms" : 10
},
{
"time_horizon" : 10000,
"number_of_arms" : 10
},
{
"time_horizon" : 1000,
"number_of_arms" : 20
},
{
"time_horizon" : 10000,
"number_of_arms" : 20
},
]
for i in range(len(params)):
results = []
time_horizon = params[i]["time_horizon"]
number_of_arms = params[i]["number_of_arms"]
agent1 = agentFactory("explore-then-exploit",time_horizon,number_of_arms,5)
agent2 = agentFactory("explore-then-exploit",time_horizon,number_of_arms,time_horizon/10)
agent3 = agentFactory("explore-then-exploit",time_horizon,number_of_arms,time_horizon/100)
bandit = Bandit(time_horizon,number_of_arms,agent1)
results.append(mc_simulate(n_sim,bandit,"N=5"))
bandit = Bandit(time_horizon,number_of_arms,agent2)
results.append(mc_simulate(n_sim,bandit,"N=T/10"))
bandit = Bandit(time_horizon,number_of_arms,agent3)
results.append(mc_simulate(n_sim,bandit,"N=T/100"))
plot(results,time_horizon,params[i])
def create_evaluation_report(
results,
k=5,
setups=None,
score='tra_score',
show_plots=False,
lower_is_better=True):
drop = list(set(results.keys()).intersection(set(nonrelevant_keys)))
results = results.drop(columns=drop)
best_results = []
# show average only if there is more than one sample
if len(np.unique(results['sample'])) > 1:
best_results.append(find_best_average(results, k=k, score=score,
lower_is_better=lower_is_better))
best_results.append(find_best(results, k=k, score=score,
lower_is_better=lower_is_better))
print("Evaluated setups: " + str(np.unique(results['setup'])))
for best in best_results:
samples = np.unique(best['sample'])
configuration_keys = list(set(results.keys()) - set(metric_keys) - set(nonrelevant_keys))
print("%d best configurations:"%k)
for sample in samples:
sample_best = best[best['sample']==sample]
first_table_keys = [
'sample',
'setup',
'iteration',
'merge_function',
'threshold',
'seg_score',
'tra_score']
table_keys = first_table_keys + [
key
for key in results.keys()
if key not in first_table_keys and key not in nonrelevant_keys
]
table_frame = sample_best.loc[:,table_keys]
if setups is not None and len(setups) > 0:
table_frame = pandas.merge(table_frame, setups, how='left', on='setup')
report.render_table(table_frame)
if show_plots:
groups = [
{
'sample': sample,
}
]
figures = [
{'x_axis': 'threshold', 'y_axis': 'seg_score', 'title': 'SEG'},
{'x_axis': 'threshold', 'y_axis': 'tra_score', 'title': 'TRA'},
]
configurations = [
# for all thresholds
dict(
{
c: curate_value(row[1][c]) for c in configuration_keys
},
**{'style':'line'}
)
for row in sample_best.iterrows()
] + [
# for best threshold only
{
c: curate_value(row[1][c]) for c in configuration_keys + ['threshold']
}
for row in sample_best.iterrows()
]
report.plot(groups, figures, configurations, results.sort_values(by='threshold'))