Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Asynchronous Distributed Gaussian Process Regression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AsyncDGP, an asynchronous distributed Gaussian process aggregation method with prediction error bounds, applied to online regression and safe control, plus the GPgym simulation platform.
Problem
Distributed Gaussian process regression for online learning suffers from asynchronous predictions and delayed measurements due to heterogeneous computation speeds across nodes, which existing aggregation methods (MOE, POE, BCM) cannot handle when inputs differ.
Approach
Aggregates all previously completed predictions from each GP expert, weighting each by a distance-aware error criterion (posterior std plus Lipschitz distance to the query state) and filtering out predictions worse than the prior, with a provable deterministic error bound and a learning-based tracking controller built on it.
Key results
- On KIN40K, SARCOS and PUMADYN32NM regression with 4 GP experts, AsyncDGP achieves lower SMSE than MOE, POE, BCM and related DGP aggregations, most clearly for information-set size 10.
- In the control simulation, AsyncDGP gives the lowest mean and median tracking and prediction errors over 100 Monte Carlo runs, with a GP error bound of 35.5 and tracking error bound of 4.9.
- The paper proves that the aggregated prediction error is bounded by beta*sigma_f (Corollary 1), independent of time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AsyncDGP instead of MOE, POE, BCM, gPOE, or rBCM | Distributed GP regression · KIN40K | SMSE | Mixed |
| Use AsyncDGP instead of MOE, POE, BCM, gPOE, or rBCM | Distributed GP regression · SARCOS | SMSE | Helped |
| Use AsyncDGP instead of MOE, POE, BCM, gPOE, or rBCM | Distributed GP regression · PUMADYN32NM | SMSE | Helped |
| Increase the information set threshold from ¯I = 4 to ¯I = 10 (ablation) | AsyncDGP · KIN40K | SMSE | Helped |
| Increase the information set threshold from ¯I = 4 to ¯I = 10 (ablation) | Distributed GP regression methods · PUMADYN32NM | SMSE | Helped |
| Increase the information set threshold from ¯I = 4 to ¯I = 10 (ablation) | Distributed GP regression methods · SARCOS | SMSE | No clear effect |
| Use AsyncDGP with ¯I = 20 for learning-based control | Dynamical system control · Control task | tracking error | Helped |
| Use AsyncDGP instead of alternative aggregation methods | Dynamical system control · Control task | prediction error | Helped |
Surprising findings
- Aggregating only the most recent available predictions is shown to be inadequate; older predictions closer to the query state remain useful.
- For SARCOS, enlarging the information set did not improve prediction accuracy, so the smaller set of 4 was better.
Limitations
- Error bounds require bounded noise and a bounded RKHS norm with a known Lipschitz kernel constant.
- Performance gains depend on the information-set size, which is not uniformly beneficial (e.g., SARCOS).
Experimental setup
Datasets: KIN40K SARCOS PUMADYN32NM
Benchmarks: KIN40K SARCOS PUMADYN32NM
Metrics: standardized mean squared error tracking error norm prediction error
Baselines: MOE POE BCM rBCM gPOE
Builds on: Gaussian process regression LoG-GP product of experts Bayesian committee machine mixture of experts
Closest papers by meaning
- Asynchronous Distributed Variational Gaussian Process for Regression (ICML 2017 · similarity 0.90)
- Streaming Generated Gaussian Process Experts for Online Learning and Control (AAAI 2026 · similarity 0.89)
- Gaussian Process-Based Real-Time Learning for Safety Critical Applications (ICML 2021 · similarity 0.87)
- Ringmaster ASGD: The First Asynchronous SGD with Optimal Time Complexity (ICML 2025 · similarity 0.84)
- Robust and Conjugate Gaussian Process Regression (ICML 2024 · similarity 0.84)
- Computation-Aware Gaussian Processes: Model Selection And Linear-Time Inference (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.