MARBLE: Multi-Armed Restless Bandits in Latent Markovian Environment

TL;DR

This paper introduces MARBLE, a model for restless bandits with latent Markovian environments, and proves convergence of a new learning algorithm under nonstationary conditions, validated on a recommender system simulator.

Contribution

It proposes MARBLE, a novel extension of RMABs with latent states, and establishes the convergence of Q-learning with Whittle Indices under relaxed indexability assumptions.

Findings

QWI adapts effectively to shifting latent states.

QWI converges to optimal policies in nonstationary environments.

MARBLE's approach is validated on a digital twin recommender system.

Abstract

Restless Multi-Armed Bandits (RMABs) are powerful models for decision-making under uncertainty, yet classical formulations typically assume fixed dynamics, an assumption often violated in nonstationary environments. We introduce MARBLE (Multi-Armed Restless Bandits in a Latent Markovian Environment), which augments RMABs with a latent Markov state that induces nonstationary behavior. In MARBLE, each arm evolves according to a latent environment state that switches over time, making policy learning substantially more challenging. We further introduce the Markov-Averaged Indexability (MAI) criterion as a relaxed indexability assumption and prove that, despite unobserved regime switches, under the MAI criterion, synchronous Q-learning with Whittle Indices (QWI) converges almost surely to the optimal Q-function and the corresponding Whittle indices. We validate MARBLE on a calibrated simulator-embedded (digital twin) recommender system, where QWI consistently adapts to a shifting latent state and converges to an optimal policy, empirically corroborating our theoretical findings.