Anytime

Read more in the user guide.

Protocol

agent.BanditAgent	Base class for anytime stochastic bandit agents.
environment.StochasticBanditEnv	Stochastic multi-armed bandit environment.
interaction.BanditInteraction	Interaction loop for the anytime stochastic bandit setting.

Agents

agents.IMED.IMED	Indexed Minimum Empirical Divergence, an asymptotically optimal bandit algorithm.
agents._Oracle.Oracle	Baseline that always plays the best arm.
agents._Random.Random	Uniform exploration: pull an arm uniformly at random every round.

Environments

Factories building a StochasticBanditEnv from a vector of arm means.

envs.parametric.BernoulliBandit	Build a Bernoulli bandit from a vector of arm means.
envs.parametric.BinomialBandit	Build a Binomial bandit from a vector of per-trial success probabilities.
envs.parametric.GaussianBandit	Build a Gaussian bandit from vectors of means and variances.
envs.parametric.TruncatedGaussianBandit	Build a bandit whose arms are Gaussians truncated to [low, high].
envs.parametric.RandomBernoulliBandit	Draw a random Bernoulli instance with a prescribed optimality gap.

Reward distributions

The individual arms the factories above are built from.

envs.distributions.Arm	One bandit arm, adapting a frozen scipy.stats distribution.
envs.distributions.Bernoulli	Bernoulli arm with success probability p, so rewards are 0 or 1.
envs.distributions.Binomial	Binomial arm: the number of successes in n trials of probability p.
envs.distributions.Gaussian	Gaussian arm of mean mu and variance var (unbounded rewards).
envs.distributions.Exponential	Exponential arm of rate p, hence mean 1 / p.
envs.distributions.TruncatedGaussian	Gaussian truncated to [low, high].
envs.distributions.TruncatedExponential	Exponential arm clipped at trunc, with mass piling on the boundary.

Renderers

renderers.textrenderer.Textrenderer

Print each bandit round to stdout, one line per pull.