Tag Archives: multiplicative weights update algorithm

The Reasonable Effectiveness of the Multiplicative Weights Update Algorithm

Posted on by
papad

Christos Papadimitriou, who studies multiplicative weights in the context of biology.

Hard to believe

Sanjeev Arora and his coauthors consider it “a basic tool [that should be] taught to all algorithms students together with divide-and-conquer, dynamic programming, and random sampling.” Christos Papadimitriou calls it “so hard to believe that it has been discovered five times and forgotten.” It has formed the basis of algorithms in machine learning, optimization, game theory, economics, biology, and more.

What mystical algorithm has such broad applications? Now that computer scientists have studied it in generality, it’s known as the Multiplicative Weights Update Algorithm (MWUA). Procedurally, the algorithm is simple. I can even describe the core idea in six lines of pseudocode. You start with a collection of $ n$ objects, and each object has a weight.

Set all the object weights to be 1.
For some large number of rounds:
   Pick an object at random proportionally to the weights
   Some event happens
   Increase the weight of the chosen object if it does well in the event
   Otherwise decrease the weight

The name “multiplicative weights” comes from how we implement the last step: if the weight of the chosen object at step $ t$ is $ w_t$ before the event, and $ G$ represents how well the object did in the event, then we’ll update the weight according to the rule:

$ \displaystyle w_{t+1} = w_t (1 + G)$

Think of this as increasing the weight by a small multiple of the object’s performance on a given round.

Here is a simple example of how it might be used. You have some money you want to invest, and you have a bunch of financial experts who are telling you what to invest in every day. So each day you pick an expert, and you follow their advice, and you either make a thousand dollars, or you lose a thousand dollars, or something in between. Then you repeat, and your goal is to figure out which expert is the most reliable.

This is how we use multiplicative weights: if we number the experts $ 1, \dots, N$, we give each expert a weight $ w_i$ which starts at 1. Then, each day we pick an expert at random (where experts with larger weights are more likely to be picked) and at the end of the day we have some gain or loss $ G$. Then we update the weight of the chosen expert by multiplying it by $ (1 + G / 1000)$. Sometimes you have enough information to update the weights of experts you didn’t choose, too. The theoretical guarantees of the algorithm say we’ll find the best expert quickly (“quickly” will be concrete later).

In fact, let’s play a game where you, dear reader, get to decide the rewards for each expert and each day. I programmed the multiplicative weights algorithm to react according to your choices. Click the image below to go to the demo.

mwua

This core mechanism of updating weights can be interpreted in many ways, and that’s part of the reason it has sprouted up all over mathematics and computer science. Just a few examples of where this has led:

  1. In game theory, weights are the “belief” of a player about the strategy of an opponent. The most famous algorithm to use this is called Fictitious Play, and others include EXP3 for minimizing regret in the so-called “adversarial bandit learning” problem.
  2. In machine learning, weights are the difficulty of a specific training example, so that higher weights mean the learning algorithm has to “try harder” to accommodate that example. The first result I’m aware of for this is the Perceptron (and similar Winnow) algorithm for learning hyperplane separators. The most famous is the AdaBoost algorithm.
  3. Analogously, in optimization, the weights are the difficulty of a specific constraint, and this technique can be used to approximately solve linear and semidefinite programs. The approximation is because MWUA only provides a solution with some error.
  4. In mathematical biology, the weights represent the fitness of individual alleles, and filtering reproductive success based on this and updating weights for successful organisms produces a mechanism very much like evolution. With modifications, it also provides a mechanism through which to understand sex in the context of evolutionary biology.
  5. The TCP protocol, which basically defined the internet, uses additive and multiplicative weight updates (which are very similar in the analysis) to manage congestion.
  6. You can get easy $ \log(n)$-approximation algorithms for many NP-hard problems, such as set cover.

Additional, more technical examples can be found in this survey of Arora et al.

In the rest of this post, we’ll implement a generic Multiplicative Weights Update Algorithm, we’ll prove it’s main theoretical guarantees, and we’ll implement a linear program solver as an example of its applicability. As usual, all of the code used in the making of this post is available in a Github repository.

The generic MWUA algorithm

Let’s start by writing down pseudocode and an implementation for the MWUA algorithm in full generality.

In general we have some set $ X$ of objects and some set $ Y$ of “event outcomes” which can be completely independent. If these sets are finite, we can write down a table $ M$ whose rows are objects, whose columns are outcomes, and whose $ i,j$ entry $ M(i,j)$ is the reward produced by object $ x_i$ when the outcome is $ y_j$. We will also write this as $ M(x, y)$ for object $ x$ and outcome $ y$. The only assumption we’ll make on the rewards is that the values $ M(x, y)$ are bounded by some small constant $ B$ (by small I mean $ B$ should not require exponentially many bits to write down as compared to the size of $ X$). In symbols, $ M(x,y) \in [0,B]$. There are minor modifications you can make to the algorithm if you want negative rewards, but for simplicity we will leave that out. Note the table $ M$ just exists for analysis, and the algorithm does not know its values. Moreover, while the values in $ M$ are static, the choice of outcome $ y$ for a given round may be nondeterministic.

The MWUA algorithm randomly chooses an object $ x \in X$ in every round, observing the outcome $ y \in Y$, and collecting the reward $ M(x,y)$ (or losing it as a penalty). The guarantee of the MWUA theorem is that the expected sum of rewards/penalties of MWUA is not much worse than if one had picked the best object (in hindsight) every single round.

Let’s describe the algorithm in notation first and build up pseudocode as we go. The input to the algorithm is the set of objects, a subroutine that observes an outcome, a black-box reward function, a learning rate parameter, and a number of rounds.

def MWUA(objects, observeOutcome, reward, learningRate, numRounds):
   ...

We define for object $ x$ a nonnegative number $ w_x$ we call a “weight.” The weights will change over time so we’ll also sub-script a weight with a round number $ t$, i.e. $ w_{x,t}$ is the weight of object $ x$ in round $ t$. Initially, all the weights are $ 1$. Then MWUA continues in rounds. We start each round by drawing an example randomly with probability proportional to the weights. Then we observe the outcome for that round and the reward for that round.

# draw: [float] -> int
# pick an index from the given list of floats proportionally
# to the size of the entry (i.e. normalize to a probability
# distribution and draw according to the probabilities).
def draw(weights):
    choice = random.uniform(0, sum(weights))
    choiceIndex = 0

    for weight in weights:
        choice -= weight
        if choice <= 0:
            return choiceIndex

        choiceIndex += 1

# MWUA: the multiplicative weights update algorithm
def MWUA(objects, observeOutcome, reward, learningRate numRounds):
   weights = [1] * len(objects)
   for t in numRounds:
      chosenObjectIndex = draw(weights)
      chosenObject = objects[chosenObjectIndex]

      outcome = observeOutcome(t, weights, chosenObject)
      thisRoundReward = reward(chosenObject, outcome)

      ...

Sampling objects in this way is the same as associating a distribution $ D_t$ to each round, where if $ S_t = \sum_{x \in X} w_{x,t}$ then the probability of drawing $ x$, which we denote $ D_t(x)$, is $ w_{x,t} / S_t$. We don’t need to keep track of this distribution in the actual run of the algorithm, but it will help us with the mathematical analysis.

Next comes the weight update step. Let’s call our learning rate variable parameter $ \varepsilon$. In round $ t$ say we have object $ x_t$ and outcome $ y_t$, then the reward is $ M(x_t, y_t)$. We update the weight of the chosen object $ x_t$ according to the formula:

$ \displaystyle w_{x_t, t} = w_{x_t} (1 + \varepsilon M(x_t, y_t) / B)$

In the more general event that you have rewards for all objects (if not, the reward-producing function can output zero), you would perform this weight update on all objects $ x \in X$. This turns into the following Python snippet, where we hide the division by $ B$ into the choice of learning rate:

# MWUA: the multiplicative weights update algorithm
def MWUA(objects, observeOutcome, reward, learningRate, numRounds):
   weights = [1] * len(objects)
   for t in numRounds:
      chosenObjectIndex = draw(weights)
      chosenObject = objects[chosenObjectIndex]

      outcome = observeOutcome(t, weights, chosenObject)
      thisRoundReward = reward(chosenObject, outcome)

      for i in range(len(weights)):
         weights[i] *= (1 + learningRate * reward(objects[i], outcome))

One of the amazing things about this algorithm is that the outcomes and rewards could be chosen adaptively by an adversary who knows everything about the MWUA algorithm (except which random numbers the algorithm generates to make its choices). This means that the rewards in round $ t$ can depend on the weights in that same round! We will exploit this when we solve linear programs later in this post.

But even in such an oppressive, exploitative environment, MWUA persists and achieves its guarantee. And now we can state that guarantee.

Theorem (from Arora et al): The cumulative reward of the MWUA algorithm is, up to constant multiplicative factors, at least the cumulative reward of the best object minus $ \log(n)$, where $ n$ is the number of objects. (Exact formula at the end of the proof)

The core of the proof, which we’ll state as a lemma, uses one of the most elegant proof techniques in all of mathematics. It’s the idea of constructing a potential function, and tracking the change in that potential function over time. Such a proof usually has the mysterious script:

  1. Define potential function, in our case $ S_t$.
  2. State what seems like trivial facts about the potential function to write $ S_{t+1}$ in terms of $ S_t$, and hence get general information about $ S_T$ for some large $ T$.
  3. Theorem is proved.
  4. Wait, what?

Clearly, coming up with a useful potential function is a difficult and prized skill.

In this proof our potential function is the sum of the weights of the objects in a given round, $ S_t = \sum_{x \in X} w_{x, t}$. Now the lemma.

Lemma: Let $ B$ be the bound on the size of the rewards, and $ 0 < \varepsilon < 1/2$ a learning parameter. Recall that $ D_t(x)$ is the probability that MWUA draws object $ x$ in round $ t$. Write the expected reward for MWUA for round $ t$ as the following (using only the definition of expected value):

$ \displaystyle R_t = \sum_{x \in X} D_t(x) M(x, y_t)$

 Then the claim of the lemma is:

$ \displaystyle S_{t+1} \leq S_t e^{\varepsilon R_t / B}$

Proof. Expand $ S_{t+1} = \sum_{x \in X} w_{x, t+1}$ using the definition of the MWUA update:

$ \displaystyle \sum_{x \in X} w_{x, t+1} = \sum_{x \in X} w_{x, t}(1 + \varepsilon M(x, y_t) / B)$

Now distribute $ w_{x, t}$ and split into two sums:

$ \displaystyle \dots = \sum_{x \in X} w_{x, t} + \frac{\varepsilon}{B} \sum_{x \in X} w_{x,t} M(x, y_t)$

Using the fact that $ D_t(x) = \frac{w_{x,t}}{S_t}$, we can replace $ w_{x,t}$ with $ D_t(x) S_t$, which allows us to get $ R_t$

$ \displaystyle \begin{aligned} \dots &= S_t + \frac{\varepsilon S_t}{B} \sum_{x \in X} D_t(x) M(x, y_t) \\ &= S_t \left ( 1 + \frac{\varepsilon R_t}{B} \right ) \end{aligned}$

And then using the fact that $ (1 + x) \leq e^x$ (Taylor series), we can bound the last expression by $ S_te^{\varepsilon R_t / B}$, as desired.

$ \square$

Now using the lemma, we can get a hold on $ S_T$ for a large $ T$, namely that

$ \displaystyle S_T \leq S_1 e^{\varepsilon \sum_{t=1}^T R_t / B}$

If $ |X| = n$ then $ S_1=n$, simplifying the above. Moreover, the sum of the weights in round $ T$ is certainly greater than any single weight, so that for every fixed object $ x \in X$,

$ \displaystyle S_T \geq w_{x,T} \leq  (1 + \varepsilon)^{\sum_t M(x, y_t) / B}$

Squeezing $ S_t$ between these two inequalities and taking logarithms (to simplify the exponents) gives

$ \displaystyle \left ( \sum_t M(x, y_t) / B \right ) \log(1+\varepsilon) \leq \log n + \frac{\varepsilon}{B} \sum_t R_t$

Multiply through by $ B$, divide by $ \varepsilon$, rearrange, and use the fact that when $ 0 < \varepsilon < 1/2$ we have $ \log(1 + \varepsilon) \geq \varepsilon – \varepsilon^2$ (Taylor series) to get

$ \displaystyle \sum_t R_t \geq \left [ \sum_t M(x, y_t) \right ] (1-\varepsilon) – \frac{B \log n}{\varepsilon}$

The bracketed term is the payoff of object $ x$, and MWUA’s payoff is at least a fraction of that minus the logarithmic term. The bound applies to any object $ x \in X$, and hence to the best one. This proves the theorem.

$ \square$

Briefly discussing the bound itself, we see that the smaller the learning rate is, the closer you eventually get to the best object, but by contrast the more the subtracted quantity $ B \log(n) / \varepsilon$ hurts you. If your target is an absolute error bound against the best performing object on average, you can do more algebra to determine how many rounds you need in terms of a fixed $ \delta$. The answer is roughly: let $ \varepsilon = O(\delta / B)$ and pick $ T = O(B^2 \log(n) / \delta^2)$. See this survey for more.

MWUA for linear programs

Now we’ll approximately solve a linear program using MWUA. Recall that a linear program is an optimization problem whose goal is to minimize (or maximize) a linear function of many variables. The objective to minimize is usually given as a dot product $ c \cdot x$, where $ c$ is a fixed vector and $ x = (x_1, x_2, \dots, x_n)$ is a vector of non-negative variables the algorithm gets to choose. The choices for $ x$ are also constrained by a set of $ m$ linear inequalities, $ A_i \cdot x \geq b_i$, where $ A_i$ is a fixed vector and $ b_i$ is a scalar for $ i = 1, \dots, m$. This is usually summarized by putting all the $ A_i$ in a matrix, $ b_i$ in a vector, as

$ x_{\textup{OPT}} = \textup{argmin}_x \{ c \cdot x \mid Ax \geq b, x \geq 0 \}$

We can further simplify the constraints by assuming we know the optimal value $ Z = c \cdot x_{\textup{OPT}}$ in advance, by doing a binary search (more on this later). So, if we ignore the hard constraint $ Ax \geq b$, the “easy feasible region” of possible $ x$’s includes $ \{ x \mid x \geq 0, c \cdot x = Z \}$.

In order to fit linear programming into the MWUA framework we have to define two things.

  1. The objects: the set of linear inequalities $ A_i \cdot x \geq b_i$.
  2. The rewards: the error of a constraint for a special input vector $ x_t$.

Number 2 is curious (why would we give a reward for error?) but it’s crucial and we’ll discuss it momentarily.

The special input $ x_t$ depends on the weights in round $ t$ (which is allowed, recall). Specifically, if the weights are $ w = (w_1, \dots, w_m)$, we ask for a vector $ x_t$ in our “easy feasible region” which satisfies

$ \displaystyle (A^T w) \cdot x_t \geq w \cdot b$

For this post we call the implementation of procuring such a vector the “oracle,” since it can be seen as the black-box problem of, given a vector $ \alpha$ and a scalar $ \beta$ and a convex region $ R$, finding a vector $ x \in R$ satisfying $ \alpha \cdot x \geq \beta$. This allows one to solve more complex optimization problems with the same technique, swapping in a new oracle as needed. Our choice of inputs, $ \alpha = A^T w, \beta = w \cdot b$, are particular to the linear programming formulation.

Two remarks on this choice of inputs. First, the vector $ A^T w$ is a weighted average of the constraints in $ A$, and $ w \cdot b$ is a weighted average of the thresholds. So this this inequality is a “weighted average” inequality (specifically, a convex combination, since the weights are nonnegative). In particular, if no such $ x$ exists, then the original linear program has no solution. Indeed, given a solution $ x^*$ to the original linear program, each constraint, say $ A_1 x^*_1 \geq b_1$, is unaffected by left-multiplication by $ w_1$.

Second, and more important to the conceptual understanding of this algorithm, the choice of rewards and the multiplicative updates ensure that easier constraints show up less prominently in the inequality by having smaller weights. That is, if we end up overly satisfying a constraint, we penalize that object for future rounds so we don’t waste our effort on it. The byproduct of MWUA—the weights—identify the hardest constraints to satisfy, and so in each round we can put a proportionate amount of effort into solving (one of) the hard constraints. This is why it makes sense to reward error; the error is a signal for where to improve, and by over-representing the hard constraints, we force MWUA’s attention on them.

At the end, our final output is an average of the $ x_t$ produced in each round, i.e. $ x^* = \frac{1}{T}\sum_t x_t$. This vector satisfies all the constraints to a roughly equal degree. We will skip the proof that this vector does what we want, but see these notes for a simple proof. We’ll spend the rest of this post implementing the scheme outlined above.

Implementing the oracle

Fix the convex region $ R = \{ c \cdot x = Z, x \geq 0 \}$ for a known optimal value $ Z$. Define $ \textup{oracle}(\alpha, \beta)$ as the problem of finding an $ x \in R$ such that $ \alpha \cdot x \geq \beta$.

For the case of this linear region $ R$, we can simply find the index $ i$ which maximizes $ \alpha_i Z / c_i$. If this value exceeds $ \beta$, we can return the vector with that value in the $ i$-th position and zeros elsewhere. Otherwise, the problem has no solution.

To prove the “no solution” part, say $ n=2$ and you have $ x = (x_1, x_2)$ a solution to $ \alpha \cdot x \geq \beta$. Then for whichever index makes $ \alpha_i Z / c_i$ bigger, say $ i=1$, you can increase $ \alpha \cdot x$ without changing $ c \cdot x = Z$ by replacing $ x_1$ with $ x_1 + (c_2/c_1)x_2$ and $ x_2$ with zero. I.e., we’re moving the solution $ x$ along the line $ c \cdot x = Z$ until it reaches a vertex of the region bounded by $ c \cdot x = Z$ and $ x \geq 0$. This must happen when all entries but one are zero. This is the same reason why optimal solutions of (generic) linear programs occur at vertices of their feasible regions.

The code for this becomes quite simple. Note we use the numpy library in the entire codebase to make linear algebra operations fast and simple to read.

def makeOracle(c, optimalValue):
    n = len(c)

    def oracle(weightedVector, weightedThreshold):
        def quantity(i):
            return weightedVector[i] * optimalValue / c[i] if c[i] &gt; 0 else -1

        biggest = max(range(n), key=quantity)
        if quantity(biggest) &lt; weightedThreshold:
            raise InfeasibleException

        return numpy.array([optimalValue / c[i] if i == biggest else 0 for i in range(n)])

    return oracle

Implementing the core solver

The core solver implements the discussion from previously, given the optimal value of the linear program as input. To avoid too many single-letter variable names, we use linearObjective instead of $ c$.

def solveGivenOptimalValue(A, b, linearObjective, optimalValue, learningRate=0.1):
    m, n = A.shape  # m equations, n variables
    oracle = makeOracle(linearObjective, optimalValue)

    def reward(i, specialVector):
        ...

    def observeOutcome(_, weights, __):
        ...

    numRounds = 1000
    weights, cumulativeReward, outcomes = MWUA(
        range(m), observeOutcome, reward, learningRate, numRounds
    )
    averageVector = sum(outcomes) / numRounds

    return averageVector

First we make the oracle, then the reward and outcome-producing functions, then we invoke the MWUA subroutine. Here are those two functions; they are closures because they need access to $ A$ and $ b$. Note that neither $ c$ nor the optimal value show up here.

    def reward(i, specialVector):
        constraint = A[i]
        threshold = b[i]
        return threshold - numpy.dot(constraint, specialVector)

    def observeOutcome(_, weights, __):
        weights = numpy.array(weights)
        weightedVector = A.transpose().dot(weights)
        weightedThreshold = weights.dot(b)
        return oracle(weightedVector, weightedThreshold)

Implementing the binary search, and an example

Finally, the top-level routine. Note that the binary search for the optimal value is sophisticated (though it could be more sophisticated). It takes a max range for the search, and invokes the optimization subroutine, moving the upper bound down if the linear program is feasible and moving the lower bound up otherwise.

def solve(A, b, linearObjective, maxRange=1000):
    optRange = [0, maxRange]

    while optRange[1] - optRange[0] &gt; 1e-8:
        proposedOpt = sum(optRange) / 2
        print(&quot;Attempting to solve with proposedOpt=%G&quot; % proposedOpt)

        # Because the binary search starts so high, it results in extreme
        # reward values that must be tempered by a slow learning rate. Exercise
        # to the reader: determine absolute bounds for the rewards, and set
        # this learning rate in a more principled fashion.
        learningRate = 1 / max(2 * proposedOpt * c for c in linearObjective)
        learningRate = min(learningRate, 0.1)

        try:
            result = solveGivenOptimalValue(A, b, linearObjective, proposedOpt, learningRate)
            optRange[1] = proposedOpt
        except InfeasibleException:
            optRange[0] = proposedOpt

    return result

Finally, a simple example:

A = numpy.array([[1, 2, 3], [0, 4, 2]])
b = numpy.array([5, 6])
c = numpy.array([1, 2, 1])

x = solve(A, b, c)
print(x)
print(c.dot(x))
print(A.dot(x) - b)

The output:

Attempting to solve with proposedOpt=500
Attempting to solve with proposedOpt=250
Attempting to solve with proposedOpt=125
Attempting to solve with proposedOpt=62.5
Attempting to solve with proposedOpt=31.25
Attempting to solve with proposedOpt=15.625
Attempting to solve with proposedOpt=7.8125
Attempting to solve with proposedOpt=3.90625
Attempting to solve with proposedOpt=1.95312
Attempting to solve with proposedOpt=2.92969
Attempting to solve with proposedOpt=3.41797
Attempting to solve with proposedOpt=3.17383
Attempting to solve with proposedOpt=3.05176
Attempting to solve with proposedOpt=2.99072
Attempting to solve with proposedOpt=3.02124
Attempting to solve with proposedOpt=3.00598
Attempting to solve with proposedOpt=2.99835
Attempting to solve with proposedOpt=3.00217
Attempting to solve with proposedOpt=3.00026
Attempting to solve with proposedOpt=2.99931
Attempting to solve with proposedOpt=2.99978
Attempting to solve with proposedOpt=3.00002
Attempting to solve with proposedOpt=2.9999
Attempting to solve with proposedOpt=2.99996
Attempting to solve with proposedOpt=2.99999
Attempting to solve with proposedOpt=3.00001
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3  # note %G rounds the printed values
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
[ 0.     0.987  1.026]
3.00000000425
[  5.20000072e-02   8.49831849e-09]

So there we have it. A fiendishly clever use of multiplicative weights for solving linear programs.

Discussion

One of the nice aspects of MWUA is it’s completely transparent. If you want to know why a decision was made, you can simply look at the weights and look at the history of rewards of the objects. There’s also a clear interpretation of what is being optimized, as the potential function used in the proof is a measure of both quality and adaptability to change. The latter is why MWUA succeeds even in adversarial settings, and why it makes sense to think about MWUA in the context of evolutionary biology.

This even makes one imagine new problems that traditional algorithms cannot solve, but which MWUA handles with grace. For example, imagine trying to solve an “online” linear program in which over time a constraint can change. MWUA can adapt to maintain its approximate solution.

The linear programming technique is known in the literature as the Plotkin-Shmoys-Tardos framework for covering and packing problems. The same ideas extend to other convex optimization problems, including semidefinite programming.

If you’ve been reading this entire post screaming “This is just gradient descent!” Then you’re right and wrong. It bears a striking resemblance to gradient descent (see this document for details about how special cases of MWUA are gradient descent by another name), but the adaptivity for the rewards makes MWUA different.

Even though so many people have been advocating for MWUA over the past decade, it’s surprising that it doesn’t show up in the general math/CS discourse on the internet or even in many algorithms courses. The Arora survey I referenced is from 2005 and the linear programming technique I demoed is originally from 1991! I took algorithms classes wherever I could, starting undergraduate in 2007, and I didn’t even hear a whisper of this technique until midway through my PhD in theoretical CS (I did, however, study fictitious play in a game theory class). I don’t have an explanation for why this is the case, except maybe that it takes more than 20 years for techniques to make it to the classroom. At the very least, this is one good reason to go to graduate school. You learn the things (and where to look for the things) which haven’t made it to classrooms yet.

Until next time!

Adversarial Bandits and the Exp3 Algorithm

Posted on by

In the last twenty years there has been a lot of research in a subfield of machine learning called Bandit Learning. The name comes from the problem of being faced with a large sequence of slot machines (once called one-armed bandits) each with a potentially different payout scheme. The problems in this field all focus on one central question:

If I have many available actions with uncertain outcomes, how should I act to maximize the quality of my results over many trials?

The deep question here is how to balance exploitation, the desire to choose an action which has payed off well in the past, with exploration, the desire to try options which may produce even better results. The ideas are general enough that it’s hard not to find applications: choosing which drug to test in a clinical study, choosing which companies to invest in, choosing which ads or news stories to display to users, and even (as Richard Feynman once wondered) how to maximize your dining enjoyment.

Herbet Robbins, one of the first to study bandit learning algorithms.

Herbert Robbins, one of the first to study bandit learning algorithms. Image credit

In less recent times (circa 1960’s), this problem was posed and considered in the case where the payoff mechanisms had a very simple structure: each slot machine is a coin flip with a different probability $ p$ of winning, and the player’s goal is to find the best machine as quickly as possible. We called this the “stochastic” setting, and last time we saw a modern strategy called UCB1 which maintained statistical estimates on the payoffs of the actions and chose the action with the highest estimate. The underlying philosophy was “optimism in the face of uncertainty,” and it gave us something provably close to optimal.

Unfortunately payoff structures are more complex than coin flips in the real world. Having “optimism” is arguably naive, especially when it comes to competitive scenarios like stock trading. Indeed the algorithm we’ll analyze in this post will take the polar opposite stance, that payoffs could conceivably operate in any manner. This is called the adversarial model, because even though the payoffs are fixed in advance of the game beginning, it can always be the case that the next choice you make results in the worst possible payoff.

One might wonder how we can hope to do anything in such a pessimistic model. As we’ll see, our notion of performing well is relative to the best single slot machine, and we will argue that this is the only reasonable notion of success. On the other hand, one might argue that real world payoffs are almost never entirely adversarial, and so we would hope that algorithms which do well theoretically in the adversarial model excel beyond their minimal guarantees in practice.

In this post we’ll explore and implement one algorithm for adversarial bandit learning, called Exp3, and in the next post we’ll see how it fares against UCB1 in some applications. Some prerequisites: since the main algorithm presented in this post is randomized, its analysis requires some familiarity with techniques and notation from probability theory. Specifically, we will assume that the reader is familiar with the content of this blog’s basic probability theory primers (1, 2), though the real difficulty in the analysis will be keeping up with all of the notation.

In case the reader is curious, Exp3 was invented in 2001 by Auer, Cesa-Bianchi, Freund, and Schapire. Here is their original paper, which contains lots of other mathematical goodies.

As usual, all of the code and data produced in the making of this blog post is available for download on this blog’s Github page.

Model Formalization and Notions of Regret

Before we describe the algorithm and analyze its we have to set up the problem formally. The first few paragraphs of our last post give a high-level picture of general bandit learning, so we won’t repeat that here. Recall, however, that we have to describe both the structure of the payoffs and how success is measured. So let’s describe the former first.

Definition: An adversarial bandit problem is a pair $ (K, \mathbf{x})$, where $ K$ represents the number of actions (henceforth indexed by $ i$), and $ \mathbf{x}$ is an infinite sequence of payoff vectors $ \mathbf{x} = \mathbf{x}(1), \mathbf{x}(2), \dots$, where $ \mathbf{x}(t) = (x_1(t), \dots, x_K(t))$ is a vector of length $ K$ and $ x_i(t) \in [0,1]$ is the reward of action $ i$ on step $ t$.

In English, the game is played in rounds (or “time steps”) indexed by $ t = 1, 2, \dots$, and the payoffs are fixed for each action and time before the game even starts. Note that we assume the reward of an action is a number in the interval $ [0,1]$, but all of our arguments in this post can be extended to payoffs in some range $ [a,b]$ by shifting by $ a$ and dividing by $ b-a$.

Let’s specify what the player (algorithm designer) knows during the course of the game. First, the value of $ K$ is given, and total number of rounds is kept secret. In each round, the player has access to the history of rewards for the actions that were chosen by the algorithm in previous rounds, but not the rewards of unchosen actions. In other words, it will only ever know one $ x_i(t)$ for each $ t$. To set up some notation, if we call $ i_1, \dots, i_t$ the list of chosen actions over $ t$ rounds, then at step $ t+1$ the player has access to the values of $ x_{i_1}(1), \dots, x_{i_t}(t)$ and must pick $ i_{t+1}$ to continue.

So to be completely clear, the game progresses as follows:

The player is given access to $ K$.
For each time step $ t$:

The player must pick an action $ i_t$.
The player observes the reward $ x_i(t) \in [0,1]$, which he may save for future use.

The problem gives no explicit limit on the amount of computation performed during each step, but in general we want it to run in polynomial time and not depend on the round number $ t$. If runtime even logarithmically depended on $ t$, then we’d have a big problem using it for high-frequency applications. For example in ad serving, Google processes on the order of $ 10^9$ ads per day; so a logarithmic dependence wouldn’t be that bad, but at some point in the distance future Google wouldn’t be able to keep up (and we all want long-term solutions to our problems).

Note that the reward vectors $ \mathbf{x}_t$ must be fixed in advance of the algorithm running, but this still allows a lot of counterintuitive things. For example, the payoffs can depend adversarially on the algorithm the player decides to use. For example, if the player chooses the stupid strategy of always picking the first action, then the adversary can just make that the worst possible action to choose. However, the rewards cannot depend on the random choices made by the player during the game.

So now let’s talk about measuring success. For an algorithm $ A$ which chooses the sequence $ i_1, \dots, i_t$ of actions, define $ G_A(t)$ to be the sum of the observed rewards

$ \displaystyle G_A(t) = \sum_{s=1}^t x_{i_s}(s)$.

And because $ A$ will often be randomized, this value is a random variable depending on the decisions made by $ A$. As such, we will often only consider the payoff up to expectation. That is, we’ll be interested in how $ \textup{E}(G_A(t))$ relates to other possible courses of action. To be completely rigorous, the randomization is not over “choices made by an algorithm,” but rather the probability distribution over sequences of actions that the algorithm induces. It’s a fine distinction but a necessary one. In other words, we could define any sequence of actions $ \mathbf{j} = (j_1, \dots, j_t)$ and define $ G_{\mathbf{j}}(t)$ analogously as above:

$ \displaystyle G_{\mathbf{j}}(t) = \sum_{s=1}^t x_{j_s}(s)$.

Any algorithm and choice of reward vectors induces a probability distribution over sequences of actions in a natural way (if you want to draw from the distribution, just run the algorithm). So instead of conditioning our probabilities and expectations on previous choices made by the algorithm, we do it over histories of actions $ i_1, \dots, i_t$.

An obvious question we might ask is: why can’t the adversary just make all the payoffs zero? (or negative!) In this event the player won’t get any reward, but he can emotionally and psychologically accept this fate. If he never stood a chance to get any reward in the first place, why should he feel bad about the inevitable result? What a truly cruel adversary wants is, at the end of the game, to show the player what he could have won, and have it far exceed what he actually won. In this way the player feels regret for not using a more sensible strategy, and likely turns to binge eating cookie dough ice cream. Or more likely he returns to the casino to lose more money. The trick that the player has up his sleeve is precisely the randomness in his choice of actions, and he can use its objectivity to partially overcome even the nastiest of adversaries.

The adversary would love to show you this bluff after you choose to fold your hand. What a jerk.

The adversary would love to show you this bluff after you choose to fold your hand. What a jerk. Image credit

Sadism aside, this thought brings us to a few mathematical notions of regret that the player algorithm may seek to minimize. The first, most obvious, and least reasonable is the worst-case regret. Given a stopping time $ T$ and a sequence of actions $ \mathbf{j} = (j_1, \dots, j_T)$, the expected regret of algorithm $ A$ with respect to $ \mathbf{j}$ is the difference $ G_{\mathbf{j}}(T) – \mathbb{E}(G_A(T))$. This notion of regret measures the regret of a player if he knew what would have happened had he played $ \mathbf{j}$.  The expected worst-case regret of $ A$ is then the maximum over all sequences $ \mathbf{j}$ of the regret of $ A$ with respect to $ \mathbf{j}$. This notion of regret seems particularly unruly, especially considering that the payoffs are adversarial, but there are techniques to reason about it.

However, the focus of this post is on a slightly easier notion of regret, called weak regret, which instead compares the results of $ A$ to the best single action over all rounds. That is, this quantity is just 

$ \displaystyle \left ( \max_{j} \sum_{t=1}^T x_j(t) \right ) – \mathbb{E}(G_A(T))$

We call the parenthetical term $ G_{\textup{max}}(T)$. This kind of regret is a bit easier to analyze, and the main theorem of this post will given an upper bound on it for Exp3. The reader who read our last post on UCB1 will wonder why we make a big distinction here just to arrive at the same definition of regret that we had in the stochastic setting. But with UCB1 the best sequence of actions to take just happened to be to play the best action over and over again. Here, the payoff difference between the best sequence of actions and the best single action can be arbitrarily large.

Exp3 and an Upper Bound on Weak Regret

We now describe at the Exp3 algorithm.

Exp3 stands for Exponential-weight algorithm for Exploration and Exploitation. It works by maintaining a list of weights for each of the actions, using these weights to decide randomly which action to take next, and increasing (decreasing) the relevant weights when a payoff is good (bad). We further introduce an egalitarianism factor $ \gamma \in [0,1]$ which tunes the desire to pick an action uniformly at random. That is, if $ \gamma = 1$, the weights have no effect on the choices at any step.

The algorithm is readily described in Python code, but we need to set up some notation used in the proof of the theorem. The pseudocode for the algorithm is as follows.

Exp3

  1. Given $ \gamma \in [0,1]$, initialize the weights $ w_i(1) = 1$ for $ i = 1, \dots, K$.
  2. In each round $ t$:
    1.  Set $ \displaystyle p_i(t) = (1-\gamma)\frac{w_i(t)}{\sum_{j=1}^K w_j(t)} + \frac{\gamma}{K}$ for each $ i$.
    2. Draw the next action $ i_t$ randomly according to the distribution of $ p_i(t)$.
    3. Observe reward $ x_{i_t}(t)$.
    4. Define the estimated reward $ \hat{x}_{i_t}(t)$ to be $ x_{i_t}(t) / p_{i_t}(t)$.
    5. Set $ \displaystyle w_{i_t}(t+1) = w_{i_t}(t) e^{\gamma \hat{x}_{i_t}(t) / K}$
    6. Set all other $ w_j(t+1) = w_j(t)$.

The choices of these particular mathematical quantities (in steps 1, 4, and 5) are a priori mysterious, but we will explain them momentarily. In the proof that follows, we will extend $ \hat{x}_{i_t}(t)$ to indices other than $ i_t$ and define those values to be zero.

The Python implementation is perhaps more legible, and implements the possibly infinite loop as a generator:

def exp3(numActions, reward, gamma):
   weights = [1.0] * numActions

   t = 0
   while True:
      probabilityDistribution = distr(weights, gamma)
      choice = draw(probabilityDistribution)
      theReward = reward(choice, t)

      estimatedReward = 1.0 * theReward / probabilityDistribution[choice]
      weights[choice] *= math.exp(estimatedReward * gamma / numActions) # important that we use estimated reward here!

      yield choice, theReward, estimatedReward, weights
      t = t + 1

Here the “rewards” variable refers to a callable which accepts as input the action chosen in round $ t$ (keeps track of $ t$, assuming we’ll play nice), and returns as output the reward for that choice. The distr and draw functions are also easily defined, with the former depending on the gamma parameter as follows:

def distr(weights, gamma=0.0):
    theSum = float(sum(weights))
    return tuple((1.0 - gamma) * (w / theSum) + (gamma / len(weights)) for w in weights)

There is one odd part of the algorithm above, and that’s the “estimated reward” $ \hat{x}_{i_t}(t) = x_{i_t}(t) / p_{i_t}(t)$. The intuitive reason to do this is to compensate for a potentially small probability of getting the observed reward. More formally, it ensures that the conditional expectation of the “estimated reward” is the actual reward. We will explore this formally during the proof of the main theorem.

As usual, the programs we write in this post are available on this blog’s Github page.

We can now state and prove the upper bound on the weak regret of Exp3. Note all logarithms are base $ e$.

Theorem: For any $ K > 0, \gamma \in (0, 1]$, and any stopping time $ T \in \mathbb{N}$

$ \displaystyle G_{\textup{max}}(T) – \mathbb{E}(G_{\textup{Exp3}}(T)) \leq (e-1)\gamma G_{\textup{max}}(T) + \frac{K \log K}{\gamma}$.

This is a purely analytical result because we don’t actually know what $ G_{\textup{max}}(T)$ is ahead of time. Also note how the factor of $ \gamma$ occurs: in the first term, having a large $ \gamma$ will result in a poor upper bound because it occurs in the numerator of that term: too much exploration means not enough exploitation. But it occurs in the denominator of the second term, meaning that not enough exploration can also produce an undesirably large regret. This theorem then provides a quantification of the tradeoff being made, although it is just an upper bound.

Proof.

We present the proof in two parts. Part 1:

We made a notable mistake in part 1, claiming that $ e^x \leq 1 + x + (e-2)x^2$ when $ x \leq 1$. In fact, this does follow from the Taylor series expansion of $ e$, but it’s not as straightforward as I made it sound. In particular, note that $ e^x = 1 + x + \frac{x^2}{2!} + \dots$, and so $ e^1 = 2 + \sum_{k=2}^\infty \frac{1}{k!}$. Using $ (e-2)$ in place of $ \frac{1}{2}$ gives

$ \displaystyle 1 + x + \left ( \sum_{k=2}^{\infty} \frac{x^2}{k!} \right )$

And since $ 0 < x \leq 1$, each term in the sum will decrease when replaced by $ \frac{x^k}{k!}$, and we’ll be left with exactly $ e^x$. In other words, this is the tightest possible quadratic upper bound on $ e^x$. Pretty neat! On to part 2:

As usual, here is the entire canvas made over the course of both videos.

$ \square$

We can get a version of this theorem that is easier to analyze by picking a suitable choice of $ \gamma$.

Corollary: Assume that $ G_{\textup{max}}(T)$ is bounded by $ g$, and that Exp3 is run with

$ \displaystyle \gamma = \min \left ( 1, \sqrt{\frac{K \log K}{(e-1)g}} \right )$

Then the weak regret of Exp3 is bounded by $ 2.63 \sqrt{g K \log K}$ for any reward vector $ \mathbf{x}$.

Proof. Simply plug $ \gamma$ in the bound in the theorem above, and note that $ 2 \sqrt{e-1} < 2.63$.

A Simple Test Against Coin Flips

Now that we’ve analyzed the theoretical guarantees of the Exp3 algorithm, let’s use our implementation above and see how it fares in practice. Our first test will use 10 coin flips (Bernoulli trials) for our actions, with the probabilities of winning (and the actual payoff vectors) defined as follows:

biases = [1.0 / k for k in range(2,12)]
rewardVector = [[1 if random.random() &lt; bias else 0 for bias in biases] for _ in range(numRounds)]
rewards = lambda choice, t: rewardVector[t][choice]

If we are to analyze the regret of Exp3 against the best action, we must compute the payoffs for all actions ahead of time, and compute which is the best. Obviously it will be the one with the largest probability of winning (the first in the list generated above), but it might not be, so we have to compute it. Specifically, it’s the following argmax:

bestAction = max(range(numActions), key=lambda action: sum([rewardVector[t][action] for t in range(numRounds)]))

Where the max function is used as “argmax” would be in mathematics.

We also have to pick a good choice of $ \gamma$, and the corollary from the previous section gives us a good guide to the optimal $ \gamma$: simply find a good upper bound on the reward of the best action, and use that. We can cheat a little here: we know the best action has a probability of 1/2 of paying out, and so the expected reward if we always did the best action is half the number of rounds. If we use, say, $ g = 2T / 3$ and compute $ \gamma$ using the formula from the corollary, this will give us a reasonable (but perhaps not perfectly correct) upper bound.

Then we just run the exp3 generator for $ T = \textup{10,000}$ rounds, and compute some statistics as we go:

bestUpperBoundEstimate = 2 * numRounds / 3
gamma = math.sqrt(numActions * math.log(numActions) / ((math.e - 1) * bestUpperBoundEstimate))
gamma = 0.07

cumulativeReward = 0
bestActionCumulativeReward = 0
weakRegret = 0

t = 0
for (choice, reward, est, weights) in exp3(numActions, rewards, gamma):
   cumulativeReward += reward
   bestActionCumulativeReward += rewardVector[t][bestAction]

   weakRegret = (bestActionCumulativeReward - cumulativeReward)
   regretBound = (math.e - 1) * gamma * bestActionCumulativeReward + (numActions * math.log(numActions)) / gamma

   t += 1
   if t &gt;= numRounds:
         break

At the end of one run of ten thousand rounds, the weights are overwhelmingly in favor of the best arm. The cumulative regret is 723, compared to the theoretical upper bound of 897. It’s not too shabby, but by tinkering with the value of $ \gamma$ we see that we can get regrets lower than 500 (when $ \gamma$ is around 0.7). Considering that the cumulative reward for the player is around 4,500 in this experiment, that means we spent only about 500 rounds out of ten thousand exploring non-optimal options (and also getting unlucky during said exploration). Not too shabby at all.

Here is a graph of a run of this experiment.

exp3-regret-graph

A run of Exp3 against Bernoulli rewards. The first graph represents the simple regret of the player algorithm against the best action; the blue line is the actual simple regret, and the green line is the theoretical O(sqrt(k log k)) upper bound. The second graph shows the weights of each action evolving over time. The blue line is the weight of the best action, while the green and red lines are the weights of the second and third best actions.

Note how the Exp3 algorithm never stops increasing its regret. This is in part because of the adversarial model; even if Exp3 finds the absolutely perfect action to take, it just can’t get over the fact that the world might try to screw it over. As long as the $ \gamma$ parameter is greater than zero, Exp3 will explore bad options just in case they turn out to be good. The benefits of this is that if the model changes over time Exp3 will adapt, but the downside is that the pessimism inherent in this worldview generally results in lower payoffs than other algorithms.

More Variations, and Future Plans

Right now we have two contesting models of how the world works: is it stochastic and independent, like the UCB1 algorithm would optimize for? Or does it follow Exp3’s world view that the payoffs are adversarial? Next time we’ll run some real-world tests to see how each fares.

But before that, we should note that there are still more models we haven’t discussed. One extremely significant model is that of contextual bandits. That is, the real world settings we care about often come with some “context” associated with each trial. Ads being displayed to users have probabilities that should take into account the information known about the user, and medical treatments should take into account past medical history. While we will not likely investigate any contextual bandit algorithms on this blog in the near future, the reader who hopes to apply this work to his or her own exploits (no pun intended) should be aware of the additional reading.

Until next time!

Postscript: years later, a cool post by Tim Vieira shows a neat data structure that asymptotically speeds up the update/sample step of the EXP3 algorithm from linear to logarithmic (among others). The weights are stored in a heap of partial sums (the leaves are the individual weights), and sampling is a binary search. See the original post and the accompanying gist for an implementation. Exercise: implement the data structure for use with our EXP3 implementation.