Physical-Risk-Aware Decentralized Agents for Constrained Vehicle Communications
Organizations: Department of Information Engineering, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong S.A.R. · State Key Laboratory of Internet of Things for Smart City and the Department of Electronic and Communication, University of Macau, Macao S.A.R.
Abstract
Decentralized V2X resource control is challenging because each vehicle has only local observations, while reliability depends on network-wide interactions. We develop a physical-risk-aware spatio-temporal graph attention (PA-STGAT) framework for agentic V2X control. Each vehicle perceives its local interaction graph, retains short-term temporal context, and independently selects its PC5 resource and power. Network outcomes are used during centralized training to refine the shared policy. Transmitter-specific credits map reliability failures to individual vehicles. A dual-priced graph term further converts network violations into coordination signals for discrete resource selection, with higher priority on short time-to-collision links. Experiments show higher raw-policy reliability than the tested graph-based baselines. Under filtered execution, PA-STGAT meets the nominal empirical communication requirements across all five seeds, with tradeoffs among utility, delivery reliability, resource occupancy, and filter-induced repair. Compared with GNN+MATD3, it requires fewer resource repairs but achieves lower utility and mean delivery reliability. Mechanism ablations show that removing dual-priced graph coordination causes resource-selection collapse. Density and mobility tests characterize the operating limits of empirical constraint satisfaction.
Figures & tables
| Symbol | Meaning |
|---|---|
| Vehicles, PC5/Uu resources, and discrete PC5 powers | |
| Communication action, raw proposal, and filtered action | |
| Selected PC5 resource/power and binary resource indicator | |
| Intended-link, successful-delivery, and low-TTC-link indicators | |
| Intended PC5 communication range | |
| Exogenous Uu grant and transmit power |
| Parameter | Value |
|---|---|
| Discount / Polyak factor | |
| Actor or Q-network / critic learning rate | |
| Batch / replay capacity | |
| Learning start / update interval | epochs |
| Dual step / cap | |
| C2 quadratic initial / growth / cap |
| Method | Raw utility | Raw PDR | Raw low-TTC | Raw C2 max | Filtered utility | Filtered PDR | Filtered low-TTC | Filtered C2 max | Filtered CBR | Resource repair | Feasible |
|---|---|---|---|---|---|---|---|---|---|---|---|
| MADQN | 5/5 | ||||||||||
| GNN+MATD3 | 5/5 | ||||||||||
| Graph+MAPPO | 5/5 | ||||||||||
| PA-STGAT | 5/5 |
| Actor | Raw utility | Raw PDR | Raw low-TTC | Raw C2 max | Filtered utility | Filtered PDR | Filtered C2 max | Filtered CBR | Feasible |
|---|---|---|---|---|---|---|---|---|---|
| PA-STGAT | 5/5 | ||||||||
| No risk relation | 4/5 | ||||||||
| Snapshot GAT | 4/5 | ||||||||
| Temporal MLP | 5/5 |
| Policy | PDR | C2 max | Resource repair | Feasible |
|---|---|---|---|---|
| AL raw | 0 | 0/5 | ||
| No-AL raw | 0 | 1.0000 | 0 | 0/5 |
| AL filtered | 5/5 | |||
| No-AL filtered | 4/5 |
| Variant | Raw PDR | Raw low-TTC | Filtered C2 max | Feasible |
|---|---|---|---|---|
| Full AL | 2/5 | |||
| No credit | 0/5 | |||
| No graph term | 0 | 0.9590 | 0/5 | |
| Generic AL | 0/5 |
| Method | Utility | PDR | CBR | C2 max | Feasible | |
| PA-STGAT | 8 | 82.03 | 0.9246 | 0.4448 | 0.0268 | 3/5 |
| 10 | 51.60 | 0.9157 | 0.5508 | 0.0200 | 5/5 | |
| 12 | 35.20 | 0.9031 | 0.6330 | 0.0247 | 4/5 | |
| 14 | 25.60 | 0.8934 | 0.7149 | 0.0293 | 0/5 | |
| 16 | 19.44 | 0.8808 | 0.7867 | 0.0419 | 0/5 | |
| MADQN | 8 | 71.20 | 0.9400 | 0.5015 | 0.0205 | 5/5 |
| Scenario | Utility | PDR | C2 max | CBR | Feasible |
|---|---|---|---|---|---|
| Nominal | 83.07 | 0.9286 | 0.0211 | 0.4392 | 5/5 |
| High speed | 80.39 | 0.9310 | 0.0589 | 0.4533 | 0/5 |
| 12 vehicles | 35.05 | 0.8983 | 0.0210 | 0.6323 | 1/5 |
| 10-ms delay | 83.06 | 0.9285 | 0.0211 | 0.4395 | 5/5 |
| 20-ms delay | 83.02 | 0.9282 | 0.0204 | 0.4394 | 5/5 |
| 10% loss | 83.09 | 0.9281 | 0.0210 | 0.4391 | 5/5 |
| Scenario | Execution | Utility | PDR | C2 max | CBR | Feasible |
|---|---|---|---|---|---|---|
| Nominal | Raw | 104.10 | 0.7507 | 0.0804 | 0.3318 | 0/5 |
| Filtered | 83.07 | 0.9286 | 0.0211 | 0.4392 | 5/5 | |
| High speed | Raw | 101.29 | 0.7297 | 0.2247 | 0.3413 | 0/5 |
| Filtered | 80.39 | 0.9310 | 0.0589 | 0.4533 | 0/5 | |
| 12 vehicles | Raw | 55.61 | 0.5765 | 0.1215 | 0.3887 | 0/5 |
| Filtered | 35.05 | 0.8983 | 0.0210 | 0.6323 | 1/5 |
V2X Communication
Vehicle-to-Everything (V2X) communication connects vehicles, infrastructure, and other road users to support safer, more efficient mobility.
See all 70 papers