This package provides tools for simulation of multi-armed bandit problems.
Pkg.add("Bandits")
There are several underlying types that need to be constructed before simulation.
The first is a Bandit
type, which specifies the true distribution of each of the arms. Currently only StaticBandit
is implemented which takes an array of Distribution
types from Distributions.jl
(i.e. staticbandit([Normal(0, 1), Uniform(0, 1)])
).
The second is a Policy
type, which specifies the policy the agent is going to follow. This type is used to specify the arm the agent should choose given the agent's beliefs over the arms.
Currently, the implemented policies are:
- Greedy
- Epsilon-Greedy
- ThompsonSampling
- UCB1
- ExploreThenExploit
The third is a Agent
type, which requires the prior of the agent, the underlying bandit, and the policy the agent should follow.
Currently, the following Agents are implemented:
- BasicAgent - this agent forms beliefs over the arms based only on observed rewards (via the empirical mean) and an initial belief about the means of the arms.
- BetaBernoulliAgent - this agent has beta priors and should be used wih Bernoulli-distributed arms. Posterior updating is done via the standard Bayesian updating formula for the Beta distribution.
- NormalAgent - this agent has Gaussian priors and should be used with Gaussian arms.
Now, we can call simulate
and get back a BanditStats
object which returns the regret and the number of times each arm was pulled.
As well, this package provides an aggregate_simulate
function which aggregates the results of N simulations run in parallel and returns the average.
Example usage as follows:
using Bandits
thompson_sampling = ThompsonSampling()
sb = staticbandit([Bernoulli(0.5), Bernoulli(0.6)])
beta_agent = BetaBernoulliAgent([0.6, 0.5], thompson_sampling, sb)
stats = simulate(sb, beta_agent, 100)
print(stats.regret)
License: MIT