An epsilon-greedy algorithm for multi-armed bandit problems
This implementation is based on Bandit Algorithms for Website Optimization and related empirical research in "Algorithms for the multi-armed bandit problem". In addition, this module conforms to the BanditLab/2.0 specification.
- Node.js 6.x+ (LTS track)
- npm
Install with npm
(or yarn
):
npm install egreedy --save
This implementation often encounters extended floating point numbers. Arm selection is therefore subject to JavaScript's floating point precision limitations. For general information about floating point issues see the floating point guide.
-
Create an optimizer with
3
arms and epsilon0.25
:const Algorithm = require('egreedy'); const algorithm = new Algorithm({ arms: 3, epsilon: 0.25 });
-
Select an arm (for exploration or exploitation, according to the algorithm):
algorithm.select().then((arm) => { // do something based on the chosen arm });
-
Report the reward earned from a chosen arm:
algorithm.reward(arm, value);
Create a new optimization algorithm.
config
(Object
): algorithm instance parameters
The config
object supports two optional parameters:
arms
(Number
, Integer): The number of arms over which the optimization will operate; defaults to2
epsilon
(Number
, Float,0
to1
): lower leads to more exploration (and less exploitation); defaults to0.5
Alternatively, the state
object resolved from Algorithm#serialize
can be passed as config
.
An instance of the egreedy optimization algorithm.
const Algorithm = require('egreedy');
const algorithm = new Algorithm();
assert.equal(algorithm.arms, 2);
assert.equal(algorithm.epsilon, 0.5);
Or, with a passed config
:
const Algorithm = require('egreedy');
const algorithm = new Algorithm({ arms: 4, epsilon: 0.75 });
assert.equal(algorithm.arms, 4);
assert.equal(algorithm.epsilon, 0.75);
Choose an arm to play, according to the specified bandit algorithm.
None
A Promise
that resolves to a Number
corresponding to the associated arm index.
const Algorithm = require('egreedy');
const algorithm = new Algorithm();
algorithm.select().then(arm => console.log(arm));
Inform the algorithm about the payoff earned from a given arm.
arm
(Number
, Integer): the arm index (provided fromAlgorithm#select()
)reward
(Number
): the observed reward value (which can be 0 to indicate no reward)
A Promise
that resolves to an updated instance of the algorithm. (The original instance is mutated as well.)
const Algorithm = require('egreedy');
const algorithm = new Algorithm();
algorithm.reward(0, 1).then(updatedAlgorithm => console.log(updatedAlgorithm));
Obtain a plain object representing the internal state of the algorithm.
None
A Promise
that resolves to a stringify-able Object
with parameters needed to reconstruct algorithm state.
const Algorithm = require('egreedy');
const algorithm = new Algorithm();
algorithm.serialize().then(state => console.log(state));
PRs are welcome! For bugs, please include a failing test which passes when your PR is applied. Travis CI provides on-demand testing for commits and pull requests.
- Feature development and bug fixing should occur on a non-master branch.
- Changes should be submitted to master via a Pull Request.
- Pull Requests should be merged via a merge commit. Local "in-process" commits may be squashed prior to pushing to the remote feature branch.
To enable a git hook that runs npm test
prior to pushing, cd
into the local repo and run:
touch .git/hooks/pre-push
chmod +x .git/hooks/pre-push
echo "npm test" > .git/hooks/pre-push
To run the unit test suite:
npm test
Or, to run the test suite and view test coverage:
npm run coverage
Note: Tests against stochastic methods (e.g. Algorithm#select
) are inherently tricky to test with deterministic assertions. The approach here is to iterate across a semi-random set of conditions to verify that each run produces valid output. As a result, each test suite run encounters slightly different execution state. In the future, the test suite should be expanded to include a more robust test of the distribution's properties – though because of the number of runs required, should be triggered with an optional flag.