24th International Conference Principles and Practice of Multi-Agent Systems, 16/11/2022


Cooperative driving at intersections through agent-based argumentation


Stefano Mariani, Dario Ferrari, Franco Zambonelli


Università di Modena e Reggio Emilia

Motivations


  • self-driving cars in the streets
  • autonomous driving $\longrightarrow$ cooperative driving [1]
  • intersection crossing pervasive problem
  • many approaches available in analysed literature [2]

[1] Englund et al.: "The grand cooperative driving challenge 2016: boosting the introduction of cooperative automated vehicles." IEEE Wireless Communications

[2] Mariani et al.: "Coordination of autonomous vehicles: Taxonomy and survey." ACM Computing Surveys

Goal & Contributions


argumentation-based coordination to complement existing approaches

  1. argumentation process
  2. proof-of-concept (PoC) simulator
  3. evaluation

Vision



  • B: "I'm going straight"
  • A: "I'm going straight, too"
  • B: "We are in conflict"
  • A: "I can't solve it, can you?"
  • B: "I could go right and reach destination anyways"
  • A: "Fine, agreed!"

Argumentation


  • "rules of the game" for human disputes resolution
  • computationally exploited for, e.g., coordination in MAS
  • in a nutshell
    • arguments as (logic) facts accepted as true until attacked
    • attacks $\approx$ conflicting facts
    • inference rules can be attacked, too
    • goal: establish valid arguments ($\approx$ not attacked)

Preference ordering


  • what if all arguments are attacked??
  • attach "strength" to arguments and inference rules to break symmetry
  • our proposal: pluggable conflicts resolution policy (e.g. alternate route)

Argumentation process


  1. incoming vehicles data $\mapsto$ argumentation graph
  2. $\forall$ vehicle check conflicts
    • no: vehicle gets right of way
    • yes: argue with conflicting vehicles
  3. conflict solved?
    • yes: give right of way accordingly
    • no: resolve with resolution policy

Evaluation setup



  1. spawn vehicles at each intersection with $A$ alternative routes of length $L$ each
  2. trigger argumentation process when vehicles within intersection range
    • $\mathtt{AltRoutes}$ = alternative routes policy
    • $\mathtt{Urgency}$ = urgency policy
    • $\mathtt{Precedence}$ as baseline
  3. update vehicles' data (e.g. pos, speed, route)

codebase: link in paper

Comparison of policies


throughput: slightly increased

delay: greatly decreased

exploiting alternative routes improves both throughput and delay w.r.t. urgency and precedence

Conflicts resolution


plotting alternative routes usage in time across networks size and routes length

  • the larger the network
  • $\rightarrow$ the more the vehicles
  • $\rightarrow$ the more the conflicts (likely)
  • $\rightarrow$ the more the alternative routes exploited (likely)

Impact on throughput


plotting throughput in time across networks size and number of alternatives

  • argumentation always better than precedence
  • the more the alternative routes $\rightarrow$ the better the throughput
  • "saturation" when $A = \frac{\text{network size}}{2}$

Impact on delay



plotting delay in time across networks size and number of alternatives

  • similar but "stronger" results than throughput
  • no "saturation" effect observed

Conclusion & outlook


  • performance improvements
  • complementary, not in competition, with State Of The Art (SOTA) strategies
  • plug-in conflicts resolution policy
  • argumentation process can be decentralised
  • naturally suitable for mixed scenarios humans / self-driving
  • naturally explainable
  • integrate with SOTA micro-traffic simulator (e.g. SUMO, MATSim)
    • realistic vehicles' dynamics
    • cost of changing route
  • translate SOTA strategies and evaluate
  • increase simulations scale
  • consider NLG for generating explanations

Thanks

for your attention


Stefano Mariani


Università di Modena e Reggio Emilia