Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Multiagent decision making for Lane-Free traffic

Troullinos Dimitrios

Full record

---

Summary

The technological and methodological progress in the field of transportation, and more specifically in vehicle automation, gives rise to research endeavours that address the existence and incorporation of Automated Vehicles (AVs) in traffic. In a future era of traffic environments without human-driven vehicles, concepts that today appear deeply-rooted might become outdated due to the enhanced observational, computational and communication capabilities of AVs when compared to human drivers. Clearly, the communication capabilities of AVs—introducing the notion of Connected and Automated Vehicles (CAVs)—unfold unprecedented possibilities for coordinated behaviour. To this end, TrafficFluid is a research project that examines the abandonment of the lane-principle in environments solely populated with CAVs. The two primary investigated aspects involve: (a) lane-free vehicle movement, meaning that CAVs are no longer restricted to specific lane-placements and can fully mobilize the available road; and (b) nudging, that is, CAVs can also react with smooth 2-dimensional movements according to the behaviour of other CAVs located in the back, thus assisting faster vehicles wishing to overtake. In the present thesis we investigate lane-free traffic with controllable CAVs under a multiagent environment. As such, each CAV is modeled as an agent, and we put forward novel approaches for multiagent decision making of CAVs in this newly-found domain. Firstly, we introduce a Factor Graph structure considering the local interactions of nearby agents that evaluates the goals regarding their individual desired speed and collision avoidance. Then, with the use of Max-Sum, a message-passing algorithm, the communicating agents exchange locally calculated messages and can then decide upon their action in a tractable way based on the intents of nearby agents as well, thereby resulting in collaborative behaviour. The local utility functions rely on a fitting ellipsoid form of Artificial Potential Fields, which effectively quantifies the risk of collision between two agents. This formulation has also resulted in extensions for single-agent Reinforcement Learning approaches and in more inclusive frameworks where external non-communicating agents occupy the road as well and introduce uncertainty to the problem.Moving forward, we focus on the message-passing algorithm and its inherent limitation regarding operation in a continuous space. In particular, the lane-free domain is a continuous one, while the Max-Sum algorithm targets discretized spaces and existing extensions towards that direction have unmet requirements for the domain of interest. To tackle this we propose the embedding of Quadtrees in Max-Sum, a tree data structure stemming from computational geometry that can approximate a 2-dimensional action space of the variables. Consequently, all locally-exchanged messages—that contain evaluation of the variable space—have a Quadtree representation that is constructed based on our approach, which results in a dynamic discretization procedure of the continuous domain. Our experimental evaluation shows that Max-Sum with Quadtrees can automatically adapt to the intensity of the problem and select additional discrete points that avoid collisions when compared to an a-priori set of specified discrete points.Then, we shift our attention towards more realistic communication frameworks, along with a reliance on other components as fallback-safety mechanisms and underlying control of CAVs. To accomplish this, we present a hybrid method consisting of a rule-based mechanism that we devise, and an updated Factor Graph approach that is distributed and considers more realistic (and limited) communication among agents based on broadcasting. The instrumental element of this approach is the introduction of dynamic lateral regions, a novel technique that allows the agents to obtain a structured representation of the surrounding lane-free traffic environment. We utilize this representation by employing models stemming from the literature for the control and regulation of agents according to safety rules. As such, the agents’ control variable, that the updated Factor Graph formulation coordinates, is now a strategic and high-level action, that of the desired lateral alignment—i.e., how can the agents be placed laterally in coordination so as to collaboratively enable faster vehicles to advance in a timely manner. The realistic communication framework for the Factor Graph is akin to the asynchrony of agents’ variable updates, while the messages calculation of Max-Sum assumes a synchronized update for the agents. Therefore, we provide Conditional Max-Sum as an extension of Max-Sum with a revised equation for the messages’ calculation, considering the relative timing of the connected agents’ upcoming updates. Experimental evaluation is conducted in both static and large-scale environments, where the proposed hybrid formulation showcases high adeptness to intense conditions, and Conditional Max-Sum exhibits higher efficiency regarding speed and comfort of agents when compared to the standard message-passing method.Finally, we present TrafficFluid-Sim, a lane-free microscopic simulator that extends the well-established SUMO simulation infrastructure to model lane-free traffic environments, allowing vehicles to be located at any lateral position, disregarding standard notions of car-following and lane-change maneuvers that are typically embedded within a (lane-based) simulator. A dynamic library has been designed for traffic monitoring and lane-free vehicle movement control, one that does not impose any inter-tool communication overhead that standard practices introduce; and enables the emulation of vehicle-to-vehicle and vehicle-to-infrastructure communication. Outside of the simulation needs on the above-mentioned research endeavours, TrafficFluid-Sim is also utilized in more complex lane-free simulation environments and emerging applications that are unique to lane-free traffic.In summary, the main contributions of this thesis are the following. Initially, we propose a formulation for Collaborative multiagent decision making in lane-free traffic, then we put forward Max-Sum with Quadtrees for decentralized coordination in continuous domains. Following that, we devise a method for Lane-free vehicle movement with dynamic lateral regions and examine the possibilities for Asynchronous decision making with the Conditional Max-Sum algorithm. Lastly, we develop a tool for Lane-free microscopic simulation of connected and automated vehicles.

Available Files

Services