LEADS: Learning Dynamical Systems that Generalize Across Environments

Yuan Yin,Ibrahim Ayed,Emmanuel de Bézenac,Nicolas Baskiotis,Patrick Gallinari

DOI: https://doi.org/10.48550/arXiv.2106.04546

2022-02-14

Abstract:When modeling dynamical systems from real-world data samples, the distribution of data often changes according to the environment in which they are captured, and the dynamics of the system itself vary from one environment to another. Generalizing across environments thus challenges the conventional frameworks. The classical settings suggest either considering data as i.i.d. and learning a single model to cover all situations or learning environment-specific models. Both are sub-optimal: the former disregards the discrepancies between environments leading to biased solutions, while the latter does not exploit their potential commonalities and is prone to scarcity problems. We propose LEADS, a novel framework that leverages the commonalities and discrepancies among known environments to improve model generalization. This is achieved with a tailored training formulation aiming at capturing common dynamics within a shared model while additional terms capture environment-specific dynamics. We ground our approach in theory, exhibiting a decrease in sample complexity with our approach and corroborate these results empirically, instantiating it for linear dynamics. Moreover, we concretize this framework for neural networks and evaluate it experimentally on representative families of nonlinear dynamics. We show that this new setting can exploit knowledge extracted from environment-dependent data and improves generalization for both known and novel environments. Code is available at <a class="link-external link-https" href="https://github.com/yuan-yin/LEADS" rel="external noopener nofollow">this https URL</a>.

Machine Learning,Artificial Intelligence

What problem does this paper attempt to address?

This paper attempts to solve the problem of how to improve the generalization ability of models in dynamic system modeling in different environments. Specifically, when modeling dynamic systems from real - world data samples, the data distribution usually changes according to the environment in which these data are captured, and the dynamic characteristics of the system also vary with different environments. Therefore, generalizing across different environments poses a challenge to traditional frameworks. The paper proposes LEADS (Learning Dynamical Systems that Generalize Across Environments), a new framework that improves the generalization ability of models by exploiting the commonalities and differences between known environments. This framework aims to capture common dynamic features through a shared model, while also capturing environment - specific dynamic features through additional terms. The paper not only theoretically proves that this method can reduce sample complexity, but also experimentally demonstrates the consistency between theory and practice in the case of simplified linear dynamics. In addition, the paper applies this framework to neural networks and experimentally evaluates it on representative families of nonlinear dynamics. The results show that this new setup can utilize the knowledge extracted from environment - dependent data to improve the generalization ability for known and new environments.

LEADS: Learning Dynamical Systems that Generalize Across Environments

Generalizing Graph ODE for Learning Complex System Dynamics across Environments

Stochastic Neural Simulator for Generalizing Dynamical Systems Across Environments

ED2: an Environment Dynamics Decomposition Framework for World Model Construction

Learning Parsimonious Dynamics for Generalization in Reinforcement Learning

Deep Active Learning by Leveraging Training Dynamics

Analyzing and Improving Supervised Nonlinear Dynamical Probabilistic Latent Variable Model for Inferential Sensors

Environment-Aware Dynamic Graph Learning for Out-of-Distribution Generalization

Curriculum learning for data-driven modeling of dynamical systems

LEAD: Exploring Logit Space Evolution for Model Selection

General Robot Dynamics Learning and Gen2Real

A Relational Intervention Approach for Unsupervised Dynamics Generalization in Model-Based Reinforcement Learning

Understanding Dynamics of Nonlinear Representation Learning and Its Application

Interpretable learning of effective dynamics for multiscale systems

Hierarchical Prototypes for Unsupervised Dynamics Generalization in Model-Based Reinforcement Learning

Prototypical context-aware dynamics generalization for high-dimensional model-based reinforcement learning

LeARN: Learnable and Adaptive Representations for Nonlinear Dynamics in System Identification

Embedding generalization within the learning dynamics: An approach based-on sample path large deviation theory

OnsagerNet: Learning Stable and Interpretable Dynamics Using a Generalized Onsager Principle

ED2: Environment Dynamics Decomposition World Models for Continuous Control

Learning Deep Hidden Nonlinear Dynamics from Aggregate Data.