Content area
Abstract
In this study, we proposed the notion of scenario generalization—a macroscopic view on the performance of an abstract decentralized crowd model, by referring a “scenario” to a configuration of environmental obstacles as well as a set of navigation tasks, specified by task attributes such as initial positions and destination positions, for all involved agents. The macroscopic input of the abstract decentralized crowd model is a scenario, and the macroscopic output is a tuple of interacting trajectories.
Scenario generalization is subtly but essentially different from the standard notion of generalization of a model. For a standard generalization, the predicted output is compared with a ground truth output, while in scenario generalization, the error is measured within the output itself (the interacting trajectories), with metrics such as the number of agent–agent collisions. For this reason, scenario generalization surprisingly shares a similar formulation with the generalization of unsupervised machine learning models. In contrast, the objective of unsupervised learning is, in general, to recover the underlying true data distribution from data samples, while in data-driven decentralized crowd modeling, a good scenario generalization implies a steering model that behaves well in an unseen target scenario domain, measured by application-oriented metrics such as the number of agent–agent collisions, without requiring that the distribution of a target scenario domain is the same as the distribution of a source scenario domain.
Based on the notion of scenario generalization, firstly, we aim to answer the question of how a training paradigm and training domain (source) affect the scenario generalization of an imitation learning model when applied to a different test domain (target). We evaluate the exact scenario generalizations of models built on combinations of imitation learning paradigms and source domains. Our empirical results suggest that (i) behavior cloning (BC) is better than generative adversarial imitation learning (GAIL); and (ii) training samples in source domain with diverse agent–agent and agent–obstacle interactions are beneficial for reducing collisions when generalized to new scenarios.
Secondly, we note that although the exact evaluation of scenario generalization is accurate, it requires training and evaluation on large datasets, coupled with complex model selection and parameter tuning. To circumvent this challenge by estimating scenario generalization without training, we proposed an information-theoretic approach to characterize both the source and target domains. The approach estimates the Interaction Score (IS) that captures the task-level inter-agent interaction difficulty of the target scenario domain. When augmented with Diversity Quantification (DQ) on the source, the combined Interaction Score Diversity Quantification (ISDQ) offers a means to estimate the source-to-target generalization of potential models. Systematic experiments verify the efficacy of ISDQ in estimating scenario generalization, compared with the exact scenario generalizations of models trained with imitation leaning paradigms (BC and GAIL) and a reinforcement learning paradigm (proximal policy optimization—PPO). Thus, ISDQ enables the rapid selection of the best source-target domain pair among multiple possible choices prior to the training and testing of the actual crowd model.
Lastly, we apply the data-driven crowd models to interpolate crowd trajectories. The goal is to fill in the gaps between observed trajectories with trajectories that contain no collisions, while reducing computational complexity. For this, we put forth an approach that consists of a data-driven prior and an optimization framework. The data-driven prior, included by crowd models, implicitly encodes the movement dependencies among neighboring agents and thus eliminates the costly pairwise collision constraints, yielding decoupled individual agent trajectories and thereby reducing computational complexity while maintaining realistic estimation. Systematically selected diverse combinations of priors and optimization algorithms are evaluated. Our experimental results reveal the essential role of these priors ensuring low-complexity, high-accuracy multi-agent trajectory interpolation.