Human Behavior Analysis via Generative Adversarial Imitation Learning

Zhang, Xin

Etd

Human Behavior Analysis via Generative Adversarial Imitation Learning

Pubblico Deposited

Technology advancement in mobile sensing and communication has enabled a massive amount of mobility data to be generated and consumed in urban life, such as GPS trajec- tories from taxis and shared bikes, working traces from gig-economy services like Door- dash and TaskRabbit, etc. These data are referred to as human-generated spatial-temporal data (HSTD). Applying the HSTD to extract the unique decision-making strategies of hu- man agents and design human-centered urban intelligent systems (e.g., self-driving ride services) has transformative potential. They can not only promote the individual well- being of gig-workers, and improve service quality and revenue of transportation service providers, but also enable downstream applications in smart transit planning, efficient gig-work dispatching, safe autonomous vehicle (AV) routing, and so on. However, analyzing human decision strategies from HSTD is a challenging task. Hu- man behaviors are complex and vary in different geographical locations over time (i.e., practical challenge), and the quality of the learned strategies is also dependent upon the model expressibility (i.e., theoretical challenge). This dissertation presents a picture of my work on human behavior analysis from HSTD based on imitation learning. They focus on tackling the above challenges by con- centrating on the following topics: Topic 1: Imitation Learning for Human Behavior Analysis. This topic identifies several practical challenges that arise when applying imitation learning to understand human behaviors in an urban scenario. Specifically, it provides solutions to spatial heterogeneity and sparsity, temporal dependency, and data quality variation when inferring human decision strategies. These practical challenges includes spatial heterogeneity and sparsity, temporal dependency, and data quality variation when inferring human decision strategy. To tackle these challenges, we designed cGAIL, TrajGAIL, and NEXT-GAIL, which are novel in algorithm design and model architecture, and are effective in providing more precise human decision understanding. Topic 2: Human Behavior Analysis Technique Improvement. This topic targets the theoretical limitations of imitation learning on human behavior analysis. The state-of- the-art approach for human behavior analysis is effective, but it relies on a predefined di- vergence measure, namely the Jensen-Shannon divergence, which limits its performance in different imitation learning tasks. To overcome this limitation, we design f-GAIL, a data-driven approach that is able to automatically search for an appropriate divergence measure to improve imitation learning performance. Moreover, we propose iMA-IL to tackle the challenging problem of multi-agent interactions in imitation learning, which is especially relevant in the field of AV and traffic coordination.

Creator