leeyangbupt / bayseanbandit Goto Github PK

Simple baysean bandit. A bandit problem involes a series of choices which
produce rewards with various probabilities. The goal is to make choices in a 
way that maximizes the total possible reward. An effective algorithm needs to 
balance between trying choices that may be better than the current best choice
(exploration) and trying the best choice found so far (exploitation). Baysean
bandit algorithms keep a calculated distribution of likely rewards for each 
choice, sample those distributions to find the best choice each round, and 
update the distribution as results are obtained. They are one of the most 
efficient algorithms, especially for the common case of the Bernoulli Bandit 
(either a reward appears or it doesn't, with constant probablity per bandit.)
This code saplies a set of bandits a given number of times and then plots 
statistics of the results.

Most code to learn bandit algorithms is done on simulated data, because 
real data is hard to obtain in practice. This code follows that lead

This code was run on Python 2.7. It requires the NumPy and matplotlib packages

The design is based on the classic Thompson Sampling algorithm: 
http://engineering.richrelevance.com/recommendations-thompson-sampling/
It tracks the number of wins and losses per bandit. These are then used to 
calculate a probability distribution of the likelyhood of each bandit paying
off, which is then sampled. In this case, the distribution is the Beta Bernouli
distribution. The bandit with the highest probability sample gets pulled each 
round. These are tracked and plotted on a scatter plot at the end of the 
sampling.

The efficiency of the algorithm is tracked by a quantity called the total 
regret: the average reward lost by picking a given bandit over the best bandit.
This is also plotted over the trials to create the regret curve. Ideally, it
will flatten out as sampling proceeds, the faster the better.

In practice, both the bandit scatter plot and the regret curve show that the
algoithm is very efficient at finding the best bandit when it has a clear
advantage over others in terms of reward rate. When two or more are close
together, the algorithm quickly selects those but can't choose between them.
Since the difference small, total regret grows slowly in this case which is 
still acceptable.

  Copyright (C) 2016   Ezra Erb

  This program is free software: you can redistribute it and/or modify
  it under the terms of the GNU General Public License version 3 as published
  by the Free Software Foundation.

  This program is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  GNU General Public License for more details.

  You should have received a copy of the GNU General Public License
  along with this program.  If not, see <http://www.gnu.org/licenses/>.

  I'd appreciate a note if you find this program useful or make
  updates. Please contact me through LinkedIn or github (my profile also has
  a link to the code depository)