AI and Optimization Group | Research

Pi Star AI and Optimization Lab


Our research studies the synthesis of reinforcement learning, convex optimization, and automated planning with applications such as robotics and intelligent transportation systems.



Deep Reinforcement Learning

What: Safely train an autonomous controller through interactions with the environment. The training process is expected to be safe, fast, and efficient. For the autonomous driving domain, we rely solely on video cameras as the controller's input.

Why: Our main goal is to foster viable autonomous transportation technology which is expected to result in a more sustainable urban society. Specifically, reduced accidents and fatalities, reduced traffic congestion, reduced CO2 emissions, increased road capacity, lower fuel consumption, increased transportation accessibility, reduced travel times, and reduced transportation costs.

How: We develop and apply state-of-the-art deep reinforcement learning algorithms for training an autonomous controller. By doing so we bypass the need to define and learn specific features from the input image. Instead, relevant features are automatically inferred through an appropriate reward function.

Who: Guni Sharon, Sheelabhadra Dey,



A demo of our Joint Imitation-Reinforcement Learning Framework (JIRL) for Safety-Critical Domains


Autonomous Intersection Controll

What: Coordinate vehicles passing through an intersection in a safe and efficient way.

Why: Travel time studies in urban areas show that 12-55% of commute travel time is due to delays induced by intersections. Hence, optimized intersection management has the potential of reducing commute time, traffic congestion, emissions, and fuel consumption, while requiring minimal infrastructure changes.

How: We develop self-learning controllers for traditional traffic signals. Such controllers utilize deep reinforcement learning algorithms to learn an optimized policy. In a different thread, we rely on the fine and accurate control of connected and autonomous vehicles along with communication capabilities to develop protocols that coordinate such vehicles to cross a given intersection in a safe and efficient way.

Who: Guni Sharon, James Ault, Aaron Parks-Young




Multiagent Pathfinding

What: In the multiagent pathfinding problem we are given a graph and a set of agents, each agent must be assigned a path leading from its initial location to its destination such that it will not collide with obstacles or other moving agents.

Why: Multiagent pathfinding is relevant to many real-world applications such as video games, traffic control, robotics, aviation, automated warehouses and more.

How: We consider various AI approaches for dealing with the combinatorial state space of this problem. Specifically, we utilize techniques from the heuristic search, multiagent planning, real-time planning, and deep-learning disciplines.

Who: Guni Sharon



Warehouse management application that is modeled as multiagent pathfinding.


Traffic Flow Optimization

What: We investigate methods by which drivers can be encouraged to follow routes that, collectively, will lead to the most efficient road usage.

Why: Growing populations in many metropolitan areas have led to noticeable rises in traffic congestion. Left to their own devices, self-interested drivers reach a user equilibrium that is far worse for the system than the ideal centralized solution that maximizes social welfare.

How: We examine game-theoretical approaches for modeling drivers' behavior and route choice. We combine various AI and convex optimization techniques to compute the optimal route assignment and the level of incentive required in order to induce optimized system performance.

Who: Guni Sharon



Delta-tolling, a practical method for computing marginal-cost pricing in traffic networks.