Anti-Jamming Task Scheduling in MEC-O-RAN With Hierarchical DRL and Transformer-Based Control

This paper presents a Deep Hierarchical Reinforcement Learning (DHRL) framework for reliable task scheduling in MEC-enabled 5G Open RAN systems under on-off jamming attacks. The scheduling problem is modeled as a combinatorial integer nonlinear program (ComINP), which is hard to solve directly. To handle this, we use Deep Reinforcement Learning (DRL). Since the jamming environment is only partially observable, the problem becomes a Partially Observable Markov Decision Process (POMDP). To deal with this, we split the problem into two DRL agents. The first agent is a jamming estimator that uses an Alternating Discrete Phase-Type Renewal Process (ADPHRP) to predict jammed time slots based on dense distribution patterns. It is trained using a Proximal Policy Optimization (PPO) algorithm. The second agent is a task scheduler called Weighted MAC-based Task Scheduler (WMAC-TS), which schedules tasks during non-jammed slots while maintaining quality of service (QoS). It uses a transformer-based Actor-Critic model with linear complexity relative to the number of tasks, considering both short-term and long-term rewards. Simulation results show that the PPO-based jamming estimator achieves a cumulative prediction error of 13 time slots in 100 time slots, compared to 25 time slots for DDQN with historical data and 48 time slots for standard DDQN. For 50 active users, WMAC-TS achieves a task drop ratio of 0.917 lower than the 0.942 of the baseline genetic algorithm, and cuts execution time from 1260 seconds to 316 seconds.