Information Directed Sampling Revisited

A random note

Posted by Ziyu Ye on May 14, 2021

This is an additional note to my previous note. Also see the new paper list on IDS, and the note on Andreas Krause’s IMSI talk.


1. High-level Idea

Decoupling Exploitation and Exploration

\begin{equation} \pi_{t}^{I D S}=\underset{\pi \in \mathcal{D}(\mathcal{A})}{\operatorname{argmin}} {\Psi_{t}(\pi):=\frac{\Delta_{t}(\pi)^{2}}{I_{t}(\pi)}\} \end{equation}

As a decision-making policy, Information-Directed Sampling (IDS) is featured by its decoupling of exploitation and exploration in optimization:

  • Exploitation is governed by immediate regret $\Delta_{t}(\pi)$.
  • Exploration is governed by mutual information (i.e., information gain) $I_{t}(\pi)$.

Exploration: Decoupling Epistemic and Aleatoric Uncertainty

Specifically in exploration, IDS is more efficient than other approaches (e.g., UCB, Thompson Sampling). The reason is that, IDS only concerns about the informative uncertainty:

  • Epistemic uncertainty: $P\left(\mathcal{E} \in \cdot \mid H_{t}\right)$, i.e., uncertainty of the environment given your observed history.
  • Aleatoric uncertainty: $P\left(O_{t+1}=\cdot \mid \mathcal{E}, H_{t}, A_{t}\right)$, i.e., uncertainty of the next observation conditioned on the environment.
  • Notice that we consider only epistemic uncertainty as the informative part of uncertainty, as it allows us to learn about the environment.

Mutual Information: Two Perspectives

As summarized in the paper list, the two lines of work describes the exploration phase of IDS by interpreting mutual information differently:

  • Van Roy’s line: $I_{t}(\mathbf{x})$ represents the reduction of uncertainty on the posterior distribution of best arm $P\left(a^{*}=a \mid \mathcal{F}_{t-1}\right)$.
  • Krause’s line: $I_{t}(\mathbf{x})$ represents the ratio of epistemic and aleatoric uncertainty. The key idea here is that not uncertainty $\neq$ informativeness, i.e., only the environment-related uncertainty is informative.

Either perspective, the key spirit is the same: we should get only information relevant to the environment.

2. Problem Statement

Setting

  • Action set: $\mathcal{A}$.
    • An agent chooses an arm $a_{t}$ at time $t \in[1, T]$.
  • Reward distribution: $p_{a}$.
    • Reward $r_{a}$ are drawn from the reward distribution $p_{a}$ of the arm $a$, and the estimated reward distribution at each time $t$ is denoted as $\hat{p}_{a,t}$.
  • Optimal arm: $a^{*}=\arg \max {a \in \mathcal{A}} \mathbb{E}{r_{a} \sim p_{a}}\left[r_{a}\right]$.
  • Belief (posterior) on best arm: $\alpha_{t}(a)=P\left(a^{*}=a \mid \mathcal{F}_{t-1}\right)$.
    • $\mathcal{F}_{t-1}$ is the past history including chosen arms and their observed rewards.
  • Policy: $\pi$.
    • This would be constructed based on the posterior $\alpha_{t}$.

Regret

\begin{equation} \mathbb{E}[\operatorname{Regret}(T)]=\underset{r_{a}^{} \sim p_{a^{}}}{\mathbb{E}} \sum_{t=1}^{T} r_{a^{*}}-\underset{a \sim \pi \atop r_{a, t} \sim p_{a}}{\mathbb{E}} \sum_{t=1}^{T} r_{a, t} \end{equation}

Immediate Regret $\Delta_{t}(a)$

\(\Delta_{t}(a)=\underset{a^{*} \sim \alpha_{t} \atop r_{a^{*}, t} \sim \hat{p}_{a^{*}, t}}{\mathbb{E}}\left[r_{a^{*}, t} \mid \mathcal{F}_{t-1}\right]-\underset{r_{a, t} \sim \hat{p}_{a, t}}{\mathbb{E}}\left[r_{a, t} \mid \mathcal{F}_{t-1}\right]\)

Information Gain $I_{t}(x)$

\(I_{t}(a) =I\left(a_{t}^{*}, r_{a, t}\right) =\underset{r_{i} \sim \hat{p}_{a,t}}\mathbb{E}\left[H\left(a_{t}^{*}\right)-H\left(a_{t+1}^{*}\right) \mid \mathcal{F}_{t-1}, a_{t}=a, r_{a, t}=r_{i}\right]\)

Optimization

\(\pi_{t}^{I D S}=\arg \min _{\pi \in \mathcal{D}(\mathcal{A})}\left\{\Psi_{t}(\pi):=\frac{\Delta_{t}(\pi)^{2}}{I_{t}(\pi)}\right\}\)

3. Key Theoretical Conclusions

Fix a deterministic $\lambda \in \mathbb{R}$ and a policy $\pi=\left(\pi_{1}, \pi_{2}, \ldots\right)$ such that $\Psi_{t}\left(\pi_{t}\right) \leq \lambda$ almost surely for each $t \in{1, . ., T} .$ Then:

\begin{aligned} \mathbb{E}(\operatorname{Regret}(T, \pi)) &=\mathbb{E} \sum_{t=1}^{T} \Delta_{t}(\pi)
& \leq \sqrt{\lambda} \mathbb{E} \sum_{t=1}^{T} \sqrt{g_{t}(\pi)}
& \leq \sqrt{\lambda T} \sqrt{\mathbb{E} \sum_{t=1}^{T} g_{t}(\pi)} \ \ \text { (Caushy-Schwardsz inequality) }
& \leq \sqrt{\lambda H\left(\alpha_{1}\right) T}. \end{aligned}

4. Connections to Combinatorial Bandits

Suppose $\mathcal{A} \subset{a \subset{0,1, \ldots, d}:|a| \leq m}$, and that there are random variables $\left(X_{t, i}: t \in \mathbb{N}, i \in{1, \ldots, d}\right)$ such that \begin{equation} Y_{t, a}=\left(X_{t, i}: i \in a\right) \quad \text { and } \quad R_{t, a}=\frac{1}{m} \sum_{i \in a} X_{t, i}. \end{equation}

Assume that the random variables $\left{X_{t, i}: i \in{1, \ldots, d}\right}$ are independent conditioned on $\mathcal{F}{t}$ and $X{t, i} \in\left[\frac{-1}{2}, \frac{1}{2}\right]$ almost surely for each $(t, i) .$ Then for all $t \in \mathbb{N}, \Psi_{t}\left(\pi_{t}^{\mathrm{IDS}}\right) \leq \frac{d}{2 m^{2}}$ almost surely. Thus:

\begin{equation} \mathbb{E}[\operatorname{Regret}(T, \pi)] \leq \sqrt{\frac{d}{2 m^{2}} H\left(\alpha_{1}\right) T}. \end{equation}

We could further prove that the lower bound of this problem is of order $\sqrt{\frac{d}{m} T}$, so the bound is order optimal to a $\sqrt{\log \left(\frac{d}{m}\right)}$ factor.

5. Can We Do Better: Adaptive IDS?

(To be updated.) An extension could be exploration inside each inner loop.