Joint Opportunistic Scheduling in Multi-Cellular Systems

We address the problem of multiuser scheduling with partial channel information in a multi-cell environment. The scheduling problem is formulated jointly with the ARQ based channel learning process and the intercell interference mitigating cell breathing protocol. The optimal joint scheduling policy under various system constraints is established. The general problem is posed as a generalized Restless Multiarmed Bandit process and the notion of indexability is studied. We conjecture, with numerical support, that the multicell multiuser scheduling problem is indexable and obtain a partial structure of the index policy.

💡 Research Summary

The paper tackles the challenging problem of multi‑user scheduling in a multi‑cellular wireless network when only partial channel state information (CSI) is available. Traditional opportunistic scheduling schemes assume perfect CSI at the scheduler, which is unrealistic because acquiring CSI incurs non‑trivial overhead. The authors therefore propose a joint design that couples (i) an inter‑cell interference (ICI) mitigation technique called “cell breathing” with (ii) an ARQ‑based channel learning mechanism that exploits ACK/NACK feedback already present for error control.

Cell breathing works by alternating the transmit power levels of neighboring cells in a rhythmic fashion: a “far” user (located farther from its serving base station) is served with high power (P1) while the adjacent cell simultaneously serves a “near” user with low power (P2 < P1). In the next time slot the roles are swapped. This power‑alternation equalizes capture probabilities for near and far users, thereby improving fairness and overall spectral efficiency without requiring extra frequency bands or complex coordination. The paper adopts this scheme as the baseline ICI control mechanism.

The channel for each user is modeled as a two‑state Gilbert‑Elliott (GE) Markov chain (ON/OFF) with transition matrix (\begin{bmatrix}p & 1-p\ r & 1-r\end{bmatrix}), where (p\ge r) captures positive temporal correlation. When a packet is transmitted, the receiver sends an ACK if the channel is ON, otherwise a NACK. The ACK/NACK is broadcast to neighboring base stations at the end of each slot, so each cell has up‑to‑date information about the channel states of all users in both cells.

The scheduling decision is made at each control interval (a time slot) by a central controller formed by the pair of base stations. Exactly one user per cell is scheduled, respecting the cell‑breathing rule (far users only when the neighboring cell serves a near user, and vice‑versa). The objective is to maximize the long‑term average sum throughput of the two‑cell system.

Mathematically, the problem is cast as a Markov Decision Process (MDP) with a very large state space: the state comprises the ARQ feedback history (i.e., the belief that each user’s channel is ON) and the current power‑level assignment dictated by the breathing pattern. Direct dynamic programming is infeasible for realistic system sizes. To obtain a tractable solution, the authors reformulate the problem as a generalized Restless Multi‑armed Bandit (RMAB) process. Each user corresponds to an “arm” that evolves (restlessly) whether or not it is scheduled, and the reward for pulling an arm is the instantaneous throughput (1 if the channel is ON, 0 otherwise).

A key concept in RMAB theory is indexability: if a system is indexable, a scalar Whittle index can be computed for each arm, and a simple index‑based policy (schedule the arms with the highest indices) is provably near‑optimal. The authors conduct an indexability analysis for the two‑cell opportunistic scheduling problem. By extensive numerical experiments they conjecture that the problem is indeed indexable, and they derive a partial analytical structure of the Whittle index. In particular, when the power levels are limited to two values (P1 and P2) and the breathing pattern is strictly followed, the index reduces to a monotone function of the belief that the user’s channel is ON. Consequently, the index‑based policy coincides with the greedy policy that maximizes the immediate expected reward.

The paper further explores three scenarios:

1. Asymmetric cooperation – the two cells have different power constraints or different channel statistics. The index is adjusted by a cell‑specific weight, and the numerical results show that the index policy still outperforms naive policies.

2. Restricted breathing patterns – the breathing sequence (e.g., Near‑Far‑Near…) must obey a predetermined periodicity. The authors incorporate this constraint into the index calculation by adding a penalty term, yielding a “constrained index” that respects the pattern while still providing a near‑optimal schedule.

3. Full cooperation – the two base stations are treated as a single controller with a joint action space. Here the state space explodes, making exact DP impossible. The RMAB formulation again offers a tractable approximation: each arm’s index is computed independently, and the joint action is simply the pair of highest‑index arms that also satisfy the breathing rule.

Simulation studies compare four policies: (i) static power with random scheduling, (ii) static power with greedy scheduling, (iii) cell‑breathing with greedy scheduling, and (iv) cell‑breathing with the proposed index‑based scheduling. Results indicate that the combination of cell breathing and ARQ‑based learning yields a 15‑30 % throughput gain over static schemes. Moreover, when the channel correlation is strong (p ≫ r), the greedy and index policies become virtually indistinguishable, confirming the analytical claim that the index reduces to the immediate reward in that regime.

In summary, the contributions of the paper are threefold:

• Introduction of a realistic, low‑overhead joint channel acquisition and scheduling framework for multi‑cell systems, built on cell breathing and ARQ feedback.
• Formalization of the joint scheduling problem as a generalized RMAB and demonstration (via numerical evidence) of its indexability.
• Derivation of a partial closed‑form structure for the Whittle index and extensive performance evaluation under various cooperation and constraint settings.

The work bridges the gap between theoretical optimal scheduling (which is computationally intractable) and practical, implementable algorithms that exploit existing control channels and power‑control mechanisms. Future research directions suggested include extending the analysis to multi‑carrier (OFDMA) systems, incorporating multiple antennas (MIMO), handling heterogeneous traffic classes, and developing low‑complexity online algorithms for real‑time deployment.