LLM-MADDPG for Multi-UAV Trajectory Design and Task Offloading
Abstract
Unmanned Aerial Vehicles (UAVs) equipped with mobile edge computing (MEC) servers can extend computation offloading services to user equipments (UEs) in areas where terrestrial infrastructure is scarce. Multi-Agent Deep Deterministic Policy Gradient (MADDPG) is a standard algorithmic choice for coordinating multiple UAVs under its centralized-training decentralized-execution (CTDE) paradigm; however, it suffers from reward sparsity and lacks an inter-agent communication mechanism in the standard formulation. We present LLM-MADDPG, which enriches MADDPG with two modules obtained from a brief offline interaction with a large language model (LLM). The first module, LLM-Guided Reward Shaping, queries the LLM with a structured description of the UAV–MEC environment to elicit domain-specific intrinsic reward factors with calibrated weights, turning the sparse environmental signal into a denser training objective. The second module, LLM-Driven Communication Protocol, uses LLM reasoning to define a 16-dimensional message schema partitioned into semantically anchored slots and learnable latent dimensions, realized by a trainable encoder–decoder pair. Both outputs are collected during a short offline phase (two API calls, no per-step LLM access) and remain fixed throughout training. On a multi-UAV MEC simulator with ITU-R P.1411 LoS/NLoS channel models, LLM-MADDPG reaches an 82.5% task completion rate—32.3 points above MADDPG and 7.9 above LMAC—while improving QoE from 1.08 to 2.31.