Multi-armed-Bandit-Algorithms

Multi-armed Bandit Algorithms in Golang.

What is this

This is the algorithm to solve multi-armed bandit problem.

https://en.wikipedia.org/wiki/Multi-armed_bandit

Though there are a lot of algorithms to solve this problem, I implemented UCB1 and epsilon-greedy strategy one so far.

To be simple, I suppose that all of success probabilities of arms will follow a bernoulli distribution.

How to Compile

$ go build main.go

How to Use

$ ./main -s=[FilePath] -i=[Number]

• -s =
  • file path for prob.d
• -i =
  • 0 : Epsilon Greedy,
  • 1 : UCB1,
  • 2 : Softmax,
  • 3 : Thompson Sampling

e.g, $ ./main -s=./arms/prob.d -i=1

You can also modify the number of trials and the maximum number of arms. (They are defined in main.go as constant.)

If you want to add another arm, you need to modify "./arms/prob.d" and add success probability of the arm.

Result (when UCB algorithm is used)

------------------------------
Total Trials:  99
Selected arm: 1
Reward: 1
[{Succecc probability, Number of selected}] =  [{0.2 18} {0.4 28} {0.8 36} {0.1 18}]
------------------------------

Arms	Success Probability	Selected Num
Arm A	0.2	18
Arm B	0.4	28
Arm C	0.8	36
Arm D	0.1	18

Comments

Using a pseudo-random number generator by leesper
https://github.com/leesper/go_rng