Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Monte carlo tree search for autonomous driving in lane-free traffic settings

Giankoulidis Pantelis

Full record

---

Summary

Lane-Free traffic is a novel paradigm that lifts the notion of lanes in traffic environments populated (as a first step, only) with autonomous vehicles, resulting in much higher efficiency since the road capacity is better exploited.Despite its novelty, a significant amount of research on lane-free autonomous driving has been already performed.However, well-known Artificial Intelligence (AI) intelligent search and decision planning algorithms have not been yet explored in this setting.To this end, we expand upon the research in lane-free vehicle movement strategies by introducing a different approach to the problem. Monte Carlo Tree Search (MCTS), a popular search algorithm for decision planning in games, is adopted for the problem at hand.We introduce a formulation for the task of lane-free driving using this algorithm and examine its efficiency under two different settings, which differ with respect to the existence of communication among vehicles, and the very constituents of the basic MCTS algorithm.In the first setting, each vehicle acts independently from the others according on the formulation of the lane-free environment that is suitable for the MCTS algorithm.The formulation we introduce addresses the two objectives of the vehicles, namely collision avoidance and reaching or preserving a desired speed of choice.While this approach gives satisfactory results, it does not take into consideration online interactions among vehicles. For every vehicle, other vehicles are observed as moving obstacles with constant speed.If we consider communication among vehicles as well, we can additionally model these interactions and examine their influence in their decision making.As such, in the second setting we address the multiagent nature of the lane-free environment. For that, we adopt a recently introduced multiagent planning algorithm based on MCTS and Coordination Graphs. Essentially, vehicles exchange messages that affect their planning, consequently resulting in a coordinated decision making process.Then, we provide an extensive set of experiments in order to evaluate our proposed approach under these two settings. Our experimental procedure correspondingly involves two distinct phases. First, the single-agent performance is investigated in scenarios populated with a large number of vehicles in a highway, all employing the MCTS algorithm independently. We observe the performance by evaluating collision occurrences and deviation from the desired speed in demanding scenarios that exceed the capacity of equivalent lane-based environments.Finally, we evaluate the multiagent algorithm in three lane-free scenarios that involve a few vehicles, and showcase its increased coordination capabilities compared to the plain MCTS algorithm at the expense of computational time.

Available Files

Services