Schedule

Keynote 1: Making AI Trustworthy for Autonomous Vehicles

Prof. Ragunathan "Raj" Rajkumar, CMU

The benefits of autonomous vehicles are yet to come to fruition in terms of safety, economic productivity, convenience and improvements in the quality of living for those unable to drive. Costly ups and downs have unfortunately been numerous along that promised pathway to a revolution in transportation. This talk will highlight the challenges and opportunities for AI in transportation while offering the view that the problem and solutions need to take a macroscopic view that is even larger than AI.   The presentation will discuss how Carnegie Mellon University tackles the essential safety, cost and scalability necessities for deploying autonomous consumer passenger vehicles.

 Contributed Talk

Peihong Yu, Manav Mishra, Syed Zaidi, Pratap Tokekar, "TACTIC: Task-Agnostic Contrastive pre-Training for Inter-Agent Communication"

Contributed Talks

Johann Schmidt, Frank Dreyer, Sayed Abid Hashimi, Sebastian Stober, "TransferLight: Zero-Shot Traffic Signal Control on any Road-Network"

Yaron Veksler, Sharon Hornstein, Han Wang, Maria Laura Delle Monache, Daniel Urieli, "Cooperative Cruising: Reinforcement Learning based Time-Headway Control for Increased Traffic Efficiency"

Anirudh Satheesh, Keenan Powell, "A Constrained Multi-Agent Reinforcement Learning Approach to Traffic Management"

Keynote 2: Model-Based Transfer Learning for Contextual Multi-Agent Traffic

Prof. Cathy Wu, MIT

Abstract: Critical to designing future traffic systems is solving numerous hard optimization and control problems. As such, researchers increasingly look to data-driven methods such as deep reinforcement learning (RL). However, deep RL often struggles with brittleness when faced with minor environmental changes. To enhance generalization across tasks, we propose Model-Based Transfer Learning (MBTL), which selects training tasks strategically. MBTL is motivated by the success of zero-shot transfer, where pre-trained models perform well on related tasks. MBTL models generalization performance using Gaussian processes for performance set points and a linear function for performance loss based on contextual similarity, combined within a Bayesian Optimization framework. Theoretically, MBTL exhibits sublinear regret in task selection. Experimental results indicate up to 50x improved sample efficiency compared to traditional training methods, including on road traffic and standard continuous control benchmarks. By layering on top of existing RL methods, MBTL paves the way to transform unreliable RL methods into more reliable ones.

Contributed Talk

 Umer Siddique, Peilang Li, Yongcan Cao, "Fairness in Traffic Control: Decentralized Multi-agent Reinforcement Learning with Generalized Gini Welfare Functions"

Contributed Talk

Patrick Benjamin, Alessandro Abate, "Addressing the MARL Scalability Problem in Autonomous Transport using Practical Mean-Field Games"

Lightning Talks

Benat Froemming-Aldanondo, Maria Gini, "A Decentralized Approach to Autonomous Train Platooning"

Joon Moon, Qadeer Ahmed, "Dynamic Shortest Path Planning with Lookahead Traffic Information"

Ashmita Bhattacharya, Malyaban Bal, "Multi-Agent Decision S4: Leveraging State Space Models for Offline Multi-Agent Reinforcement Learning"