Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2402.01467

Brain-Like Replay Naturally Emerges in Reinforcement Learning Agents

Jiyi Wang·Likai Tang·Huimiao Chen·Marcelo G Mattar·Sen Song
View original source ↗Read source document ↗Export metadata ↓

Abstract

Replay is a powerful strategy to promote learning in artificial intelligence and the brain. However, the conditions to generate it and its functional advantages have not been fully recognized. In this study, we develop a modular reinforcement learning model that could generate replay. We prove that replay generated in this way helps complete the task. We also analyze the information contained in the representation and provide a mechanism for how replay makes a difference. Our design avoids complex assumptions and enables replay to emerge naturally within a task-optimized paradigm. Our model also reproduces key phenomena observed in biological agents. This research explores the structural biases in modular ANN to generate replay and its potential utility in developing efficient RL.

Explore related subjects

eess.SYcs.AIcs.CEcs.NEcs.SYq-bio.NC

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jiyi Wang, Likai Tang, Huimiao Chen, Marcelo G Mattar, Sen Song. 2024-10-06. Brain-Like Replay Naturally Emerges in Reinforcement Learning Agents. https://arxiv.org/abs/2402.01467

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Observability and parameter estimation of a generic model for aggregated distributed energy resources

We propose a novel framework for estimating the parameters of an aggregated distributed energy resources (DER A) model. First, we introduce a rigorous method to determine whether all model parameters are estimable. When they are not, our approach identifies the subset of parameters that can be estimated. The proposed framework offers new insights into the number and specific parameters that can be reliably estimated based on commonly available measurements. It also highlights the limitations of calibrating such models. Second, we introduce a Kalman filtering method to calibrate the DER A model. Since we account for nonlinear effects such as saturation and deadbands, we develop a specific mechanism to handle smoothing functions within the Kalman filter. Specifically, we consider the extended and the unscented Kalman filter. We demonstrate the effectiveness of the proposed framework on a modified IEEE 34-node distribution feeder with inverter- based resources. Our findings align with the North American Electric Reliability Corporation's parameterization guideline and underscore the importance of model calibration in accurately capturing the collective dynamics of distributed energy resources installed on distribution systems.

eess.SY↗
arXiv2026

Salted Fisher Information for Hybrid Systems

Discrete events change how parameter-influence propagates in hybrid systems. Prevailing Fisher information for- mulations assume that sensitivities evolve smoothly according to continuous-time variational equations and therefore neglect the sensitivity updates induced by discrete events. This paper derives a Fisher information matrix formulation compatible with hybrid systems. To do so, we use the saltation matrix, which encodes the first-order transformation of sensitivities induced by discrete events. We call the resulting formulation the salted Fisher information matrix (SFIM). The proposed framework unifies continuous information accumulation during flows with discrete updates at event times. We also show that hybrid persistence of excitation is sufficient for the SFIM to be positive definite

eess.SY↗
arXiv2026

Min-Max Grassmannian Optimization for Online Subspace Tracking

We propose GeRoST (Geometrically Robust Subspace Tracking), an online subspace tracking algorithm that models uncertainty in a subspace using a Grassmannian ball. We derive an exact scalar dual for the worst-case subspace problem, establish conditions for a unique worst-case subspace and a Riemannian gradient, and characterize the minimum radius needed to cover a dimensional extension of the target subspace. Each update uses either a spectral direction computed in a reduced subspace or the gradient of the window reconstruction loss. Our numerical experiments show that GeRoST achieves lower mean post-fault prediction error than GREAT in system identification. In video separation, it achieves higher precision and a better precision--recall balance, as measured by the F$_1$ score, than both GREAT and GRASTA at the reported thresholds, with lower recall and longer runtime.

eess.SY↗