def history(means, algorithm, seed, horizon, epsilon=0.02):
'''
Returns regret history from T = 0 to T = horizon
'''
np.random.seed(seed)
bandit = BanditInstance(means)
optimal_arm = np.amax(means)
reward_sum = 0
regret = np.zeros(horizon + 1)
strategy = 0
if algorithm == 'epsilon-greedy':
strategy = EpsilonGreedy(means.shape[0], epsilon)
elif algorithm == 'ucb':
strategy = UCB(means.shape[0])
elif algorithm == 'kl-ucb':
strategy = KLUCB(means.shape[0])
elif algorithm == 'thompson-sampling':
strategy = ThompsonSampling(means.shape[0])
elif algorithm == 'thompson-sampling-with-hint':
strategy = ThompsonSamplingWithHint(means.shape[0], np.sort(means))
for i in range(1, horizon + 1):
arm = strategy.getArm()
reward = bandit.pull(arm)
strategy.getReward(arm, reward)
reward_sum += reward
regret[i] = i * optimal_arm - reward_sum
return regret
experiment = Experiment(1)
experiment.log_code()
# Experiment 1
N = 50
epsilon = .3
simulations = 10000
T = 400
algorithms = [
GeneralCausal(truncate='None'),
ParallelCausal(),
SuccessiveRejects(),
AlphaUCB(2),
ThompsonSampling()
]
m_vals = range(2, N, 2)
regret, models = regret_vs_m(algorithms,
m_vals,
N,
T,
epsilon,
simulations=simulations)
experiment.plot_regret(regret,
m_vals,
"m",
algorithms,
legend_loc="lower right")
return m_vals,regret,models
experiment = Experiment(4)
experiment.log_code()
N = 50
N1_vals = range(1,N,3)
pz = .4
q = (0.00001,0.00001,.4,.65)
epsilon = .3
simulations = 10000
T = 400
algorithms = [SuccessiveRejects(),GeneralCausal(),AlphaUCB(2),ThompsonSampling()]
epsilon = .3
pY = ParallelConfounded.pY_epsilon_best(q,pz,epsilon)
m_vals,regret,models = regret_vs_m_general(algorithms,N1_vals,N,T,pz,pY,q,epsilon,simulations = simulations)
experiment.plot_regret(regret,m_vals,"m",algorithms,legend_loc = "lower right",legend_extra = [ParallelCausal])