Event Details


PhD Defense - Nitin Kamra

Mon, May 10, 2021
10:30 AM - 12:00 PM
PhD Candidate: Nitin Kamra

Committee: Prof. Yan Liu (chair), Prof. Bistra Dilkina, Prof. Ashutosh Nayyar

Date: 10th May, 2021
Time: 10:30am-12:00pm

Zoom: https://usc.zoom.us/j/96990524233?pwd=M3BzcTdOenZtRjlkN1J5dmxDQmVpUT09

Meeting ID: 969 9052 4233
Passcode: 015551

Title: Machine Learning in Interacting Multi-agent Systems


Making predictions and learning optimal behavioral strategies are important problems in many domains such as traffic prediction, pedestrian tracking, financial investments and security systems. These systems often consist of multiple agents interacting with each other in complex ways, which makes both the above tasks very challenging. In this thesis, I propose methods to advance the state-of-the-art for such multi-agent learning problems. The first part of my talk focuses on trajectory prediction and I will present a relational model involving a fuzzy decision making attention mechanism for multi-agent trajectory prediction. Our approach shows significant performance gains over many existing state-of-the-art predictive models in diverse domains such as human crowds, US freeway traffic and various physics datasets. The second part of my talk focuses on placing multiple resources to protect and cover geographical spaces. We propose the Coverage Gradient Theorem and a spatial discretization based framework to improve existing benchmarks for spatial coverage domains. The third part of the talk focuses on computing nash equilibrium strategies in spatial security games with continuous action spaces. We present our model-free learning algorithm, OptGradFP, and our model-based learning algorithm, DeepFP, which search for the optimal defender strategy in a parameterized continuous search space. These algorithms scale to large domains and compute strategies robust to adversarial exploitation. Finally, we combine the Coverage Gradient framework with DeepFP to show improved performance on spatial coverage security domains.