Tactical Decision-Making for Highway Driving

This thesis investigates three different Monte Carlo tree search (MCTS) algorithms for optimizing tactical decision-making during highway driving. The optimization problem was expressed in a partially observable Markov decision process (POMDP) framework, where the behaviors of the surrounding vehicles were modeled as nonobservable variables. The motion of the vehicles were governed by a generative model, which used two conventional driver models; the intelligent driver model (IDM) and minimizing overall braking induced by lane changes (MOBIL). These models together contain eight parameters for each vehicle which estimate a vehicle’s behaviour with respect to its longitudinal motion and lane changes. These eight non-observable parameters were inferred by a particle filter. The algorithms were tested in a simulated environment, where the objective was to change lanes to reach an exit ramp in dense highway traffic. The results show that the partially observable Monte Carlo planning (POMCP) based algorithms require more computational effort to reach the same performance as the MCTS based one, due to the inherent complexity of the history node trees. However, both methods are feasible to implement as an online tactical decision-making algorithm, where the less complex MCTS method performs best during simulations with limited resources