Upper Confidence Bounds for Multi-Arm Bandits
Sutton and Barto do a great job of explaining many things but the derivation of Upper Confidence Bounds is not great.
Let’s get to whether they got differently.
Problems with \(\epsilon\)-greedy algorithms
An \(\epsilon\)-greedy solution selects a random action with probability \(\epsilon\) and selects the greedy action
the the rest of the time.
If we assume that the means for the different “arms” are stationary, i.e. not changing over time, \(\epsilon\)-greedy algorithms will always, even after we’ve been training for eons, select every action with probability \(\frac{\epsilon}{k}\).
Ideally, we want to explore a lot at the beginning and exploit more as we sample more from the different arms. An \(\epsilon\)-greedy algorithm makes no distinction between when we have 100000 samples or just 1 sample to decide how much or how to explore.
We could try to manually decay \(\epsilon\), but that feels pretty hard to get right…
Vanilla Upper Confidence Bounds
When we get more samples, we learn more about the variance of the different arms.
One way to formalize this is to construct a confidence interval and then pick the arm which has the highest upper confidence bound.
Lets say that we want to construct a \(95\)% confidence interval for \(\mathbb{E}[R_a] = \mu_a\). Recal the Central Limit Theorem allows us to find a confidence interval for the rewards from an action \(a\):
\[\overline{x} \pm \Phi^{-1}(.025)\cdot \frac{\sigma_{R_a}}{\sqrt{n}}\]This suggests the vanilla confidence bound algorithm:
- Pick the action with the highest \(95\)% upper confidence bound. \(A_{t}\doteq \underset{a}{\arg\max}\left[Q_{t}(a) + \Phi^{-1}(.975)\cdot \sigma_{R_a} \frac{1}{\sqrt{N_t(a)}}\right]\)
Note:
- \(N_t(a) = \sum\limits_{i=1}^{t-1} \mathbf{1}\{a_t = a\}\) is the number of times we selected action \(a\) before time \(t\).
- \(Q_t(a)\) is our estimate for \(\mathbb{E}[R_a]\) the mean of arm \(a\) at time \(t\).
OK, cool, but there’s one glaring problem here: We don’t know \(\sigma_{R_a}\). We could estimate it using the sample variance and use a quantile in the \(t\)-distribution instead of the normal, but what if we just, gasps, make it a hyper-parameter and make it our user’s problems? Lets call that \(c_a\)
\[\begin{align} A_{t} &\doteq \underset{a}{\arg\max}\left[Q_{t}(a) + \Phi^{-1}(.975)\cdot \sigma_{R_a} \frac{1}{\sqrt{N_t(a)}}\right] \\ &=\underset{a}{\arg\max}\left[Q_{t}(a) + $\Phi^{-1}(.975) \cdot c_a \frac{1}{\sqrt{N_t(a)}}\right] \end{align}\]Agumenting Upper Confidence Bounds for Exploration
Now, a problem with this is that you might get extraordinarily unlucky and pull a couple really low rewards for the optimal arm. If that happens, our estimate \(Q_t(a)\) of \(\mathbb{E}[R_a] = \mu_*\) will be way too low and we’ll never recover because our upper confidence bound will be far too low (recall that \(95\)% of confidence intervals will contain the true parameter, so that means \(5\)% of them don’t!).
To deal with that, one elegant way is to increase the confidence level of our confidence intervals over time (this is equivalent to keeping the confidence level the same and increasing the variance estimate) for each estimate of \(\mathbb{E}[R_a] = \mu_a\).
We can do that with a multipler which we’ll call \(\lambda_a(t)^2\).
\[A_{t} =\underset{a}{\arg\max}\left[Q_{t}(a) + c_a \cdot \sqrt{\lambda_a(t)^2} \cdot \frac{1}{\sqrt{N_t(a)}}\right]\]Well, we want \(\lambda_a(t)\) to decrease over time, right? But, we want to make sure that it’s always unbounded so that as we go to \(\infty\) in time, the confidence intervals go to \(100\)%. What if we set it to \(\lambda_a(t)^2 = \ln t\)?. This would make it so that at \(100\) samples, we’re multiplying by \(\sqrt(\ln 100) = 2.14 \approx \Phi(.99)\), so we’ll be at a \(98\)% confidence interval after \(100\) samples.
\[\sqrt{\lambda_a(t)^2} \cdot \frac{1}{\sqrt{N_t(a)}} = \sqrt{\frac{\lambda_a(t)^2}{N_t(a)}} = \sqrt{\frac{\ln t}{\ln t}} = \sqrt{1} = 1\]If all of the means are of the same magnitude, we’re encouraged to have \(O(\ln t)\) samples for each action \(a\) at any given point in time. If we have less than \(\ln t\) samples for a point in time \(t\), then the upper confidence bound for \(a\) will be more than \(c\) away from the estimated mean.
Then, we get
\[\begin{align} A_{t} &=\underset{a}{\arg\max}\left[Q_{t}(a) + c_a \cdot \sqrt{\lambda_a(t)^2} \cdot \frac{1}{\sqrt{N_t(a)}}\right] \\ &= \underset{a}{\arg\max}\left[Q_{t}(a) + c_a \cdot \sqrt{\ln t} \cdot \frac{1}{\sqrt{N_t(a)}}\right] \\ &= \underset{a}{\arg\max}\left[Q_{t}(a) + c_a \cdot \sqrt{\frac{\ln t}{N_t(a)}}\right] \end{align}\]This is the final formulation in the book by Sutton and Barto.
The Gorey Derivation of UCB
I kind of lied before when I said we’re “arbitrarily” picking $\n t$. It turns out that it comes from hoeffding’s inequality for 1 sub-gaussian variables. We’ll derive that now.