class UCBExample:
label = '2.6 - Upper-Confidence-Bound Action Selection'
bandit = GaussianBandit(10)
agents = [
Agent(bandit, GreedyPolicy(1)),
Agent(bandit, GreedyPolicy(1), prior=10),
Agent(bandit, EpsilonGreedyPolicy(0.1, 1)),
Agent(bandit, UCBPolicy(2)),
]
class OptimisticInitialValueExample:
label = 'Optimistic Initial Values'
bandit = GaussianBandit(10)
agents = [
Agent(bandit, EpsilonGreedyPolicy(0.1, 1)),
Agent(bandit, GreedyPolicy(1)),
Agent(bandit, GreedyPolicy(1), prior=1),
Agent(bandit, GreedyPolicy(1), prior=2),
Agent(bandit, GreedyPolicy(1), prior=5),
Agent(bandit, GreedyPolicy(1), prior=10),
]
class EpsilonGreedyExample:
label = 'Action-Value Methods'
bandit = GaussianBandit(10)
GreedyPolicy(1)
agents = [
Agent(bandit, GreedyPolicy(1)),
Agent(bandit, EpsilonGreedyPolicy(0.01, 1)),
Agent(bandit, EpsilonGreedyPolicy(0.1, 1)),
# Agent(bandit, EpsilonGreedyPolicy(0.1, 1), prior=5),
# Agent(bandit, GreedyPolicy(1), prior=5)
# Agent(bandit, GreedyPolicy(10)),
# Agent(bandit, EpsilonGreedyPolicy(0.1, 10)),
]
class BinomialExample:
label = 'Bayesian Bandits - Binomial (n=5)'
bandit = BinomialBandit(10, n=5, t=3 * 1000)
agents = [
Agent(bandit, EpsilonGreedyPolicy(0.1)),
Agent(bandit, UCBPolicy(1)),
BetaAgent(bandit, GreedyPolicy())
]
class BernoulliExample:
label = 'Bayesian Bandits - Bernoulli'
bandit = BernoulliBandit(10, t=3 * 1000)
agents = [
Agent(bandit, EpsilonGreedyPolicy(0.1)),
Agent(bandit, UCBPolicy(1)),
BetaAgent(bandit, GreedyPolicy())
]
class EpsilonGreedyExample:
label = '2.2 - Action-Value Methods'
bandit = GaussianBandit(10)
agents = [
Agent(bandit, GreedyPolicy()),
Agent(bandit, EpsilonGreedyPolicy(0.01)),
Agent(bandit, EpsilonGreedyPolicy(0.1)),
]
class CompareExample:
label = 'Action'
n_arms = 10
n_trials = 1000
bandits = [GaussianBandit(n_arms), GaussianBandit(n_arms), GaussianBandit(n_arms), GaussianBandit(n_arms),
GaussianBandit(n_arms, mu=4), BernoulliBandit(n_arms, t=3 * n_trials)]
agents = [
Agent(bandits[0], GreedyPolicy(1)),
Agent(bandits[1], EpsilonGreedyPolicy(0.1, 1)),
Agent(bandits[2], GreedyPolicy(1), prior=5),
Agent(bandits[3], UCBPolicy(2)),
GradientAgent(bandits[4], SoftmaxPolicy(), alpha=0.1, baseline=False),
BetaAgent(bandits[5], GreedyPolicy())
# Agent(bandit, EpsilonGreedyPolicy(0.1, 1), prior=5),
# Agent(bandit, GreedyPolicy(1), prior=5)
]
"""
Takes advantage of multicore systems to speed up the simulation runs.
"""
from bandits.agent import Agent, BetaAgent
from bandits.bandit import BernoulliBandit
from bandits.policy import GreedyPolicy, EpsilonGreedyPolicy, UCBPolicy
from bandits.environment import Environment
if __name__ == '__main__':
experiments = 500
trials = 1000
bandit = BernoulliBandit(10, t=3 * 1000)
agents = [
Agent(bandit, EpsilonGreedyPolicy(0.1)),
Agent(bandit, UCBPolicy(1)),
BetaAgent(bandit, GreedyPolicy())
]
env = Environment(bandit, agents, label='Bayesian Bandits')
scores, optimal = env.run(trials, experiments)
env.plot_results(scores, optimal)
env.plot_beliefs()