Bernoulli bandits an empirical comparison

K. N. Ronoh; R. Oyamo; E. Milgo; M. Drugan; B. Manderick

Bernoulli bandits an empirical comparison

K. N. Ronoh, R. Oyamo, E. Milgo, M. Drugan, B. Manderick

Security

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

3 Citations (Scopus)

Abstract

An empirical comparative study is made of a sample of action selection policies on a test suite of the Bernoulli multi-armed bandit with Κ = 10, Κ = 20 and Κ = 50 arms, each for which we consider several success probabilities. For such problems the rewards are either Success or Failure with unknown success rate. Our study focusses on ε-greedy, UCB1-Tuned, Thompson sampling, the Gittin's index policy, the knowledge gradient and a new hybrid algorithm. The last two are not wellknown in computer science. In this paper, we examine policy dependence on the horizon and report results which suggest that a new hybridized procedure based on Thompsons sampling improves on its regret.

Original language	English
Title of host publication	23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN 2015 - Proceedings
Publisher	i6doc.com publication
Pages	59-64
Number of pages	6
ISBN (Print)	9782875870148
Publication status	Published - 2015
Event	23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2015) - Hotel Novotel, Bruges, Belgium Duration: 22 Apr 2015 → 24 Apr 2015 Conference number: 23

Conference

Conference	23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2015)
Abbreviated title	ESANN 2015
Country/Territory	Belgium
City	Bruges
Period	22/04/15 → 24/04/15
Other

Cite this

@inproceedings{08ef787a9ed94f7f85324e3459082008,

title = "Bernoulli bandits an empirical comparison",

abstract = "An empirical comparative study is made of a sample of action selection policies on a test suite of the Bernoulli multi-armed bandit with Κ = 10, Κ = 20 and Κ = 50 arms, each for which we consider several success probabilities. For such problems the rewards are either Success or Failure with unknown success rate. Our study focusses on ε-greedy, UCB1-Tuned, Thompson sampling, the Gittin's index policy, the knowledge gradient and a new hybrid algorithm. The last two are not wellknown in computer science. In this paper, we examine policy dependence on the horizon and report results which suggest that a new hybridized procedure based on Thompsons sampling improves on its regret.",

author = "Ronoh, {K. N.} and R. Oyamo and E. Milgo and M. Drugan and B. Manderick",

year = "2015",

language = "English",

isbn = "9782875870148",

pages = "59--64",

booktitle = "23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN 2015 - Proceedings",

publisher = "i6doc.com publication",

note = "23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2015), ESANN 2015 ; Conference date: 22-04-2015 Through 24-04-2015",

}

Ronoh, KN, Oyamo, R, Milgo, E, Drugan, M & Manderick, B 2015, Bernoulli bandits an empirical comparison. in 23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN 2015 - Proceedings. i6doc.com publication, pp. 59-64, 23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2015), Bruges, Belgium, 22/04/15.

Bernoulli bandits an empirical comparison. / Ronoh, K. N.; Oyamo, R.; Milgo, E. et al.

23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN 2015 - Proceedings. i6doc.com publication, 2015. p. 59-64.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Bernoulli bandits an empirical comparison

AU - Ronoh, K. N.

AU - Oyamo, R.

AU - Milgo, E.

AU - Drugan, M.

AU - Manderick, B.

N1 - Conference code: 23

PY - 2015

Y1 - 2015

N2 - An empirical comparative study is made of a sample of action selection policies on a test suite of the Bernoulli multi-armed bandit with Κ = 10, Κ = 20 and Κ = 50 arms, each for which we consider several success probabilities. For such problems the rewards are either Success or Failure with unknown success rate. Our study focusses on ε-greedy, UCB1-Tuned, Thompson sampling, the Gittin's index policy, the knowledge gradient and a new hybrid algorithm. The last two are not wellknown in computer science. In this paper, we examine policy dependence on the horizon and report results which suggest that a new hybridized procedure based on Thompsons sampling improves on its regret.

AB - An empirical comparative study is made of a sample of action selection policies on a test suite of the Bernoulli multi-armed bandit with Κ = 10, Κ = 20 and Κ = 50 arms, each for which we consider several success probabilities. For such problems the rewards are either Success or Failure with unknown success rate. Our study focusses on ε-greedy, UCB1-Tuned, Thompson sampling, the Gittin's index policy, the knowledge gradient and a new hybrid algorithm. The last two are not wellknown in computer science. In this paper, we examine policy dependence on the horizon and report results which suggest that a new hybridized procedure based on Thompsons sampling improves on its regret.

UR - http://www.scopus.com/inward/record.url?scp=84961780776&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:84961780776

SN - 9782875870148

SP - 59

EP - 64

BT - 23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN 2015 - Proceedings

PB - i6doc.com publication

T2 - 23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2015)

Y2 - 22 April 2015 through 24 April 2015

ER -

Bernoulli bandits an empirical comparison

Abstract

Conference

Other files and links

Fingerprint

Cite this