ATutorialonThompsonSampling - Stanford University

1 A Tutorial on Thompson SamplingDaniel J. Russo1, Benjamin Van Roy2, Abbas Kazerouni2, IanOsband3and Zheng Wen41 Columbia University2 Stanford University3 Google DeepMind4 Adobe ResearchABSTRACTT hompson sampling is an algorithm for online decision prob-lems where actions are taken sequentially in a manner thatmust balance between exploiting what is known to maxi-mize immediate performance and investing to accumulatenew information that may improve future performance. Thealgorithm addresses a broad range of problems in a compu-tationally efficient manner and is therefore enjoying wideuse. This tutorial covers the algorithm and its application,illustrating concepts through a range of examples, includingBernoulli bandit problems, shortest path problems, productrecommendation, assortment, active learning with neuralnetworks, and reinforcement learning in Markov decisionprocesses. Most of these problems involve complex informa-tion structures, where information revealed by taking anaction informs beliefs about other actions.

2 We will also dis-cuss when and why Thompson sampling is or is not effectiveand relations to alternative memory of Arthur F. Veinott, multi-armed bandit problem has been the subject of decades ofintense study in statistics, operations research, electrical engineering,computer science, and economics. A one-armed bandit is a somewhatantiquated term for a slot machine, which tends to rob players of theirmoney. The colorful name for our problem comes from a motivatingstory in which a gambler enters a casino and sits down at a slot machinewith multiple levers, or arms, that can be pulled. When pulled, an armproduces a random payout drawn independently of the past. Becausethe distribution of payouts corresponding to each arm is not listed, theplayer can learn it only by experimenting. As the gambler learns aboutthe arms payouts, she faces a dilemma: in the immediate future sheexpects to earn more byexploitingarms that yielded high payouts inthe past, but by continuing toexplorealternative arms she may learnhow to earn higher payouts in the future.

3 Can she develop a sequentialstrategy for pulling arms that balances this tradeoff and maximizes thecumulative payout earned? The following Bernoulli bandit problem is acanonical (Bernoulli Bandit) Suppose there areKactions, andwhen played, any action yields either a success or a failure. Action34 Introductionk {1,..,K}produces a success with probability k [0,1]. Thesuccess probabilities( 1,.., K)are unknown to the agent, but arefixed over time, and therefore can be learned by experimentation. Theobjective, roughly speaking, is to maximize the cumulative number ofsuccesses overTperiods, whereTis relatively large compared to thenumber of arms in this problem might represent different banner adsthat can be displayed on a website. Users arriving at the site are shownversions of the website with different banner ads. A success is associatedeither with a click on the ad, or with a conversion (a sale of the itembeing advertised).

4 The parameters krepresent either the click-through-rate or conversion-rate among the population of users who frequent thesite. The website hopes to balance exploration and exploitation in orderto maximize the total number of naive approach to this problem involves allocating some fixedfraction of time periods to exploration and in each such period samplingan arm uniformly at random, while aiming to select successful actionsin other time periods. We will observe that such an approach can bequite wasteful even for the simple Bernoulli bandit problem describedabove and can fail completely for more complicated like the Bernoulli bandit described above have been studiedin the decision sciences since the second world war, as they crystallize thefundamental trade-off between exploration and exploitation in sequentialdecision making. But the information revolution has created significantnew opportunities and challenges, which have spurred a particularlyintense interest in this problem in recent years.

5 To understand this,let us contrast the Internet advertising example given above with theproblem of choosing a banner ad to display on a highway. A physicalbanner ad might be changed only once every few months, and onceposted will be seen by every individual who drives on the road. There isvalue to experimentation, but data is limited, and the cost of of tryinga potentially ineffective ad is enormous. Online, a different banner adcan be shown to each individual out of a large pool of users, and datafrom each such interaction is stored. Small-scale experiments are now acore tool at most leading Internet interest in this problem is motivated by this broad learning is increasingly used to make rapid data-driven deci-sions. While standard algorithms in supervised machine learning learnpassively from historical data, these systems often drive the generationof their own training data through interacting with users.

6 An onlinerecommendation system, for example, uses historical data to optimizecurrent recommendations, but the outcomes of these recommendationsare then fed back into the system and used to improve future recom-mendations. As a result, there is enormous potential benefit in thedesign of algorithms that not only learn from past data, but also exploresystemically to generate useful data that improves future are significant challenges in extending algorithms designed toaddress Example to treat more realistic and complicated decisionproblems. To understand some of these challenges, consider the problemof learning by experimentation to solve a shortest path (Online Shortest Path) An agent commutes from hometo work every morning. She would like to commute along the path thatrequires the least average travel time, but she is uncertain of the traveltime along different routes. How can she learn efficiently and minimizethe total travel time over a large number of trips?

7 Figure :Shortest path can formalize this as a shortest path problem on a graphG= (V,E)with verticesV={1,..,N}and edgesE. An exampleis illustrated in Figure Vertex1is the source (home) and vertexNis the destination (work). Each vertex can be thought of as an intersec-tion, and for two verticesi,j V, an edge(i,j) Eis present if thereis a direct road connecting the two intersections. Suppose that travelingalong an edgee Erequires time eon average. If these parameterswere known, the agent would select a path(e1,..,en), consisting of asequence of adjacent edges connecting vertices1andN, such that theexpected total time e1+..+ enis minimized. Instead, she chooses pathsin a sequence of periods. In periodt, the realized timeyt,eto traverseedgeeis drawn independently from a distribution with mean e. Theagent sequentially chooses a pathxt, observes the realized travel time(yt,e)e xtalong each edge in the path, and incurs costct= e xtyt,eequal to the total travel time.

8 By exploring intelligently, she hopes tominimize cumulative travel time Tt=1ctover a large number of problem is conceptually similar to the Bernoulli bandit inExample , but here the number of actions is the number of pathsin the graph, which generally scales exponentially in the number ofedges. This raises substantial challenges. For moderate sized graphs,trying each possible path would require a prohibitive number of samples,and algorithms that require enumerating and searching through theset of all paths to reach a decision will be computationally efficient approach therefore needs to leverage the statistical andcomputational structure of this model, the agent observes the travel time along each edgetraversed in a given period. Other feedback models are also natural: theagent might start a timer as she leaves home and checks it once shearrives, effectively only tracking the total travel time of the chosen is closer to the Bernoulli bandit model, where only the realizedreward (or cost) of the chosen arm was observed.

9 We have also taken therandom edge-delaysyt,eto be independent, conditioned on e. A morerealistic model might treat these as correlated random variables, reflect-ing that neighboring roads are likely to be congested at the same than design a specialized algorithm for each possible statistical7model, we seek a general approach to exploration that accommodatesflexible modeling and works for a broad array of problems. We will seethat Thompson sampling accommodates such flexible modeling, andoffers an elegant and efficient approach to exploration in a wide rangeof structured decision problems, including the shortest path problemdescribed sampling also known asposterior samplingandprobabil-ity matching was first proposed in 1933 (Thompson, 1933; Thompson,1935) for allocating experimental effort in two-armed bandit problemsarising in clinical trials. The algorithm was largely ignored in theacademic literature until recently, although it was independently re-discovered several times in the interim (Wyatt, 1997; Strens, 2000) asan effective heuristic.

10 Now, more than eight decades after it was intro-duced, Thompson sampling has seen a surge of interest among industrypractitioners and academics. This was spurred partly by two influentialarticles that displayed the algorithm s strong empirical performance(Chapelle and Li, 2011; Scott, 2010). In the subsequent five years, theliterature on Thompson sampling has grown rapidly. Adaptations ofThompson sampling have now been successfully applied in a wide vari-ety of domains, including revenue management (Ferreiraet al., 2015),marketing (Schwartzet al., 2017), web site optimization (Hillet al.,2017), Monte Carlo tree search (Baiet al., 2013), A/B testing (Graepelet al., 2010), Internet advertising (Graepelet al., 2010; Agarwal, 2013;Agarwalet al., 2014), recommendation systems (Kawaleet al., 2015),hyperparameter tuning (Kandasamyet al., 2018), and arcade games(Osbandet al., 2016a); and have been used at several companies, includ-ing Adobe, Amazon (Hillet al.)