Bandit Algorithms. Zhifeng Wang ... Department of Statistics Florida State University

Transcription

1 Bandit Algorithms Zhifeng Wang Department of Statistics Florida State University

2 Outline Multi-Armed Bandits (MAB) Exploration-First Epsilon-Greedy Softmax UCB Thompson Sampling Adversarial Bandits Exp3 Contextual Bandits LinUCB

3 K Slot Machines......

4 K Slot Machines Choose a machine and get a reward Only have T chances Goal: maximize the cumulative rewards How to choose the machines (arms)?

5 Two Phases: Explore-Exploit Explore: Try different arms. Exploit: Play the most rewarding arm. A/B Testing: Draw the arms uniformly during T/2 rounds Then draw the empirical best until the end Jump from pure exploration to pure exploitation. Bandits: Transit from exploration to exploitation more smoothly.

6 A/B Testing v.s. Bandits......

7 Multi-Armed Bandits (MAB) A set of K arms, denoted by A. The number of rounds T is fixed. At the round t, player picks arm at a it A. Player gets rewards r it [0, 1] for the chosen arm (no full information). Assume for each arm a, its rewards iid D a (unknown) People usually call this iid setting stochastic bandits.

8 Regret Regret: R(T ) = T µ T µ(a t ), t=1 where µ(a i ) := E(r i ) denotes the mean reward for the arm a i and µ := max a A µ(a) is the optimal mean reward. Maximizing rewards Minimizing regret.

9 MAB Algorithms......

10 Exploration-First Algorithm A/B testing style Exploration phase: try each arm N times Select the arm a opt with the highest average reward Exploitation phase: play arm a opt in all remaining rounds Regret bound: with a proper value of N, R(T ) 1 + 2K log(2kt ) 2, min where min = min a A\{a } ( µ µ(a) ).

11 Epsilon-Greedy Algorithm Exploit: with probability 1 ϵ, play the best arm so far Explore: with probability ϵ, choose an arm randomly

12 Epsilon-Greedy Algorithm If ϵ is held constant, only a linear bound on R(T ) can be achieved. When ϵ decreases with time, it could obtain ( ) K log T R(T ) O 2 min with ϵ t = O(1/t).

13 Adaptive Exploration A big flaw: the exploration is completely random, does not depend on the history of the observed rewards. Adaptive v.s. non-adaptive exploration.

14 Softmax Algorithm Idea: pick each arm with a probability that is proportional to its average reward. Given initial empirical means ˆµ i for i = 1,..., K P(Play kth Arm) = ˆµ k K i=1 ˆµ i Softmax Algorithm (Boltzmann Exploration): P(Play kth Arm) = exp(ˆµ k /τ) K i=1 exp(ˆµ i/τ) τ: a temperature parameter, controlling the randomness of the choice. Annealing: dynamically decrease τ over time.

15 Epsilon-Greedy and Softmax Tend to select the best arm currently. Sometimes decide to explore and choose an option that is not currently best. Drawbacks: ignore the randomness (noise) of rewards could be easily misled by negative experiences: non-robust

16 UCB1 Algorithm Given initial empirical means ˆµ i for i = 1,..., K Play the jth arm with j = arg max UCB(a i ) := ˆµ i + i=1,...,k α log t n i, where n i is the number of times a i was played so far. Usually, α = 2.

17 UCB UCB: Upper Confidence Bounds UCB(a i ) := ˆµ i + α log t n i By Hoeffding s inequality ( ) ( α log t ( α log t )) P ˆµ i + µ i exp 2n i = 1 n i n i t 2α

18 Regret Bound for UCB1 Regret Bound for UCB1 (Auer et al., 2002) E(R(T )) 8 where i = µ µ i. i: i >0 ( K log T O min No need to set any parameter! log T i + ), ) (1 + π2 K i, 3 i=1

19 UCB1-Tuned A variant which takes into account the variance of each arm. Play the jth arm with j = arg max i=1,...,k ˆµ i + log t n i ( 1 min i) 4, V, where V i = ˆσ 2 i + 2 log t ˆσ 2 i can be computed from the historical rewards. n i

20 Thompson Sampling Bayes point of view: quantify the reward in terms of a distribution rather than a point estimate. Prior: assume rewards follow Beta(S, F ). S: wins, F: fails. Set S i = F i = 0 for i = 1,..., K. At the t-th round, Sample θ i Beta(S i + 1, F i + 1), i. Play arm j t := arg max i θ i and receive reward r t Generate w t Ber(r t ). Update S jt = S jt + 1 if w t = 1, otherwise F jt = F jt + 1.

21 Regret Bound for Thompson Sampling Regret Bound for Thompson sampling (Agrawal and Goyal, 2012) ( ( 1 ) ) 2 E(R(T )) O log T. 2 i: i >0 i This rate is the same as that in UCB1, but is inferior in terms of constant factors and dependence on. Cheap in computation and have competitive performance to UCB1.

22 Comparison We generated bernoulli rewards with K = 5, T = Figure: Average of 100 runs

23 Adversarial Bandits Non-stochastic: sometimes the reward distribution cannot be modeled by a stationary distribution. Adversarial bandit game: at time t The player chooses arm a it Simultaneously, an adversary chooses a vector of rewards [r t 1, r t 2,..., r t K] The player only receives the reward r t i t. We still assume r [0, 1] for simplicity.

24 Adversarial Bandits For any deterministic algorithm there exists a sequence of rewards such that R(T ) T/2. The idea is to add randomization to the selection of the arm.

25 Exp3 Algorithm Set p 0 i = 1/K, i = 1,..., K and G 0 i = 0 (estimated cumulative rewards) At the round t (t 1): Sample arm i based on the probability p t 1, observe r t i t Estimating rewards for all actions: g t = [0,..., 0, Update cumulative rewards: Update sampling probability ri t t p t 1 i G t = G t 1 + g t, 0,..., 0] p t = exp( ηgt ) 1, exp( ηg t )

26 Exp3 Algorithm Unbiased estimate of unseen rewards [ r t E[gj] t j = E A variant: p t 1 j 1 j=it ] = rt j p t 1 j E[1 j=it ] = rt j p t 1 j p t 1 j p t exp( ηg t ) = (1 γ) 1, exp( ηg t ) + γ 1, γ (0, 1] K = r t j A mixture of the uniform distribution (exploration) and a exponential weights (exploitation).

27 Contextual Bandits Some additional information available at each round. This information (context) could help with the arm choices. Web article recommendation system. Contextual information about visitors: demographic browsing history location

28 Contextual Bandits Each round t proceeds The player observes a context x t The player chooses an arm a it Reward r t [0, 1] is realized Reward distribution could be stochastic (iid) or adversarial.

29 Example: Web Article Recommendation 3 articles but only one space: K = 3 2 user features: if they had clicked on a sports article or a politics article in the past. Find which articles are best for people given their past click behaviors. arm clk sports clk politics reward

30 LinUCB A combination of supervised learning and UCB At each round, fit a ridge regression for each arm ˆθ a = (X a X a + I) 1 X a r a X a is the feature matrix for the arm a so far. r a is the reward vector from choosing the arm a so far. [ ] 1 0 X 1 =, X = [ 0 0 ], X 3 = [ 0 1 ] r 1 = [ ] 0.58, r = [ 0.19 ], r 3 = [ 0.51 ]

31 LinUCB Choose the arm which gives the largest UCB for a new observed context x t ( ) a + = arg max x ˆθ t a +α x a A }{{} t (X a X a + I) 1 x t }{{} prediction standard deviation

32 LinUCB Feature engineering is extremely important. Could use both user features and arm features. Hybrid-LinUCB allows arms to share contextual variables. GLM-UCB for rewards following distributions from exponential family.