Sequential game solution for lane-merging conflict between autonomous vehicles

Abstract

Lane-merging conflict between Autonomous Vehicles (AV) calls for coordinated solution to allocate right-of-way. Related studies resort to centralized decision-making optimization models such as “right-of-way reservation/auction”, which are suitable only for the scenarios with a centralized intersection agent (acting as arbiter or auctioneer); and involved fiat currency spent on bidding may trigger controversial issues concerning law or taxation. This paper: (i) establishes a prototype of 2-player complete-information 3-stage sequential game architecture within distributed decision-making paradigm, to formalize the interactions between 2 AVs trapped in the lane-merging conflict, so as to suit for scenarios with or without a centralized decision-maker, i.e. intersection or road segment; (ii) designs the dynamic rewards of AV?s game-playing (velocity-adjusting) actions which result from the space-time status of AV; (iii) based on the proposed rewards, uses Multi-agent Reinforcement Learning to obtain the optimal (in Nash equilibrium sense) strategies of action-sequence for both AVs after 3-stage game-theoretic negotiations, promisingly avoiding the potential right-of-way deadlock in a lane-merging conflict.