Intersection crossing: problem

• decision making process orchestrating vehicles' actions
• necessary to achieve own goal
• competitive access to shared resource
• poses 3 challenges
• safety: no collisions
• liveness: right of way for everyone
• quality: max./min. throughput/delay

Intersection crossing: solutions

• today: traffic light, precedence rules
• safe & inefficient
• efficient & unsafe
tomorrow: thanks to V2I / V2V
• reservation
• negotiation (e.g. auction)
• others (e.g. DCOP, game-theoretic, self-org)

Selected approaches

• precedence-based right of way as baseline
centralised reservation-based as "best in class"
• IM receives space-time reservations for intersection cells from approaching vehicles
• IM elaborates collision-free trajectories
• IM communicates back to vehicles their parameters for crossing (e.g. speed profile)
decentralised competitive auction as state-of-art
• vehicles bid for space-time slots within the intersection
• non-colliding vehicles cross simultaneously
• a broker collects bids and ranks them to assign the right of way

Performance comparison

• SUMO simulations, controlled via TraCi API (www.eclipse.org/sumo/)
• reservation based on temporal allocations of the intersection area, divided into a grid of occupancy cells [1]
• English auctions with virtual wallets [2]

[1] Kurt M. Dresner, Peter Stone: A Multiagent Approach to Autonomous Intersection Management. J. Artif. Intell. Res. 31: 591-656 (2008)

[2] Dustin Carlino, Stephen D. Boyles, Peter Stone: Auction-based autonomous intersection management. ITSC 2013: 529-534

Simulation settings

• single 4-ways junction
• 3 lanes for each way
• 1 to 4 vehicles/second (traffic condition)
• 33/66 % of left turning
• 30 runs for each params combination

Results

• no approach works best across traffic conditions
• no approach works best across performance metrics

Adaptive approach

• learn which approach works best in which traffic condition
• apply the best approach known upon changing conditions
• both with respect to a target performance metric

municipalities already do this!
(e.g. traffic lights turned off at night)

Adaptive Intersection Manager (AIM)

learning phase (naive Q-learning*):
1. randomly select coordination approach $A$
2. apply $A$ for a given time $T$
3. monitor both traffic conditions $C$ and target performance metric $M$ during $T$
adaptation phase:
1. sample traffic condition $C$
2. lookup best approach $A$ for $C$ regarding $M$
3. apply $A$ until $C \rightarrow C'$
4. loop

* action = application of $a \in A$ during $T$, reward = $M$ measured during $T$

Results

• the AIM pursues the best $A$ known for every $C$
• the AIM accounts for the target $M$ to optimise