MORL-CA: An IJCAI 2026-Accepted Framework for Multi-Objective Reinforcement Learning in Cellular Network Automation

Introducing the MORL-CA framework—an open-source implementation accepted by the IJCAI 2026 AI4Tech track. This framework applies multi-objective reinforcement learning (MORL) to address complex resource optimization challenges in 5G/6G cellular networks, offering a flexible approach to balance conflicting goals like spectrum efficiency, user QoS, energy consumption, coverage, and handover performance.

Modern cellular networks face exponential complexity with 5G deployment and ongoing 6G development. Traditional rule-based optimization methods struggle to adapt to dynamic network environments. Key challenges include balancing conflicting objectives:

• Spectrum efficiency (maximize data per unit spectrum)
• User QoS (guarantee delay/bandwidth for different services)
• Energy consumption (reduce base station/network energy use)

Single-objective RL has limitations: weight selection difficulty, lack of Pareto front exploration, and poor adaptability to changing environments. Multi-objective RL directly handles vector rewards to learn Pareto-optimal strategies, providing flexible decision options for operators.

The MORL-CA framework integrates multi-objective optimization, deep RL, and network simulation. Core components:

• State Space: Multi-dimensional tensor including user device state (position, speed, service type, signal quality), base station state (load, power, available resource blocks), channel state (interference, SNR), and network topology.
• Action Space: Mixed continuous-discrete actions like power control, resource block allocation, handover decisions, carrier aggregation, and beamforming.
• Reward Function: Vector form covering throughput, delay, energy consumption, fairness index, and coverage quality—no manual weight adjustment needed.

MORL-CA implements advanced MORL algorithms:

• Decomposition-based methods (MOEA/D variants)
• Pareto-based methods (Pareto Q-Learning)
• Preference-guided methods (for priority-based scenarios)

Deep neural network optimizations: Graph neural networks for base station relationships; temporal models for traffic patterns.

The framework includes a high-fidelity simulation environment with 3GPP channel models, real topology support, traffic generators, and performance evaluation tools.

Typical application scenarios:

• Dynamic spectrum management (maximize utilization while ensuring QoS)
• Energy-saving optimization (dynamic base station power adjustment/sleep)
• Load balancing (balance base station loads via user association)

Experimental results:

• 15-30% throughput improvement over traditional algorithms