Opinion Dynamic Models of Decision-Making in Traffic

Licentiatavhandling, 2024

Autonomous vehicles (AVs) have the potential to improve both the efficiency and the safety of road traffic. Vehicles that can plan their routes, anticipate accidents and communicate open the door to transportation that is nearly free of human error. However, in the transition toward fully autonomous transportation systems, AVs must handle the dangers of human road users (HRUs).

   One of the challenges faced by AVs is to accurately interpret and predict human behavior and decision-making. Due to the vast number of factors that influence every single individual, precise, deterministic models of decision-making between humans are practically infeasible. Moreover, while AVs may exchange explicit, technical information about each other’s decisions, such communication might be difficult between them and HRUs. As a result, HRUs introduce an element of uncertainty in traffic scenes.

   While many methods for estimating HRU decision-making are based on data-driven machine-learning methods, model-based approaches that use data for calibration are advantageous for simulation and prediction due to their relatively low parameter complexity. However, such models need to describe decision-making using stochastic abstractions that also capture the effect that interaction between HRUs has on their decision-making process. In this thesis, a framework based on Markovian opinion dynamics is suggested for modelling human decision-making in traffic as a network of continuous-time Markov chain agents that randomly switch between decisions. Interaction is expressed as social forces that modulate the rates at which agents change their own decisions depending on the decisions of others. The probability of intuitive effects such as group-wise agreement and disagreement can be predicted based on the modeled interaction within and between groups of agents.

   The model can be used to anticipate how traffic scenario probabilities evolve from an initial observation to a stationary prediction. This thesis suggests how such a transition can be derived over the horizon of a predictive controller that determines the acceleration of an AV based on the expected HRU decision-making process.