Never-Overconfident Geometric Pooling of Sparse Gaussian Process Evidence Across Robot Teams

A team of robots mapping an unknown field, such as gas concentration, signal strength or temperature, faces a basic problem. Each robot sees only the strip of ground it has driven over, and the robots can talk only to nearby teammates. To decide where to go next, each robot needs an estimate of the whole field and an honest sense of how uncertain that estimate is. A Gaussian process (a probabilistic model that predicts a smooth function together with an error bar at every location) gives such estimates. The question is how robots should merge their Gaussian processes over a peer-to-peer network that has no central server.

The social-learning literature has a clear answer for discrete hypotheses: average log-beliefs with neighbours, which is called geometric pooling. Applied to Gaussian processes, this rule has a known failure. If agents multiply full predictive distributions, one agent that is confidently wrong can crush everyone's error bars. This proposal argues that the failure comes from pooling the wrong object. Agents should pool only the evidence their own data contributes, written in a compact sparse-Gaussian-process basis that all agents share. They should then rescale that pooled evidence by a factor tied to how well the network has mixed. With this rescaling, no agent's posterior is ever more confident than the posterior a central computer would compute from all the data. The posterior becomes exactly central once mixing completes. Memory stays fixed regardless of mission length or graph diameter.

Core Claim

Two facts, put together, give the proposal its mechanism.

The first fact concerns sparse Gaussian processes. In a sparse Gaussian process with $m$ inducing points $Z$ (a small set of locations where the function value summarizes the data) and a Gaussian likelihood, the optimal posterior over the inducing values $u$ has natural parameters that add up across data points. Natural parameters here means the precision matrix and the precision-weighted mean. The posterior precision is the prior precision $K_{zz}^{-1}$ plus the sum of the evidence matrices $\Lambda_j$ contributed by each agent $j$'s data. The precision-weighted mean likewise has the form $\eta = \sum_j \eta_j$.

The second fact concerns geometric pooling. In natural parameters, geometric pooling is linear averaging. This is the multiplicative identity behind both Non-Bayesian social learning and the product-of-experts aggregators.

The proposal is to gossip only the evidence pair $(\Lambda_j,\eta_j)$, never the posterior, using a doubly stochastic weight matrix $W$. After $t$ rounds, agent $i$ forms

$$ P_i^{(t)} = K_{zz}^{-1} + s_t \sum_j [W^t]_{ij}\,\Lambda_j, \qquad s_t = \frac{1}{1/N + \lambda^t}, $$

where $N$ is the number of agents, $\lambda$ is the second-largest singular value of $W$, and the precision-weighted mean $\eta$ is scaled the same way.

The scale $s_t$ comes from a standard mixing bound, $\max_j [W^t]_{ij} \le 1/N + \lambda^t$. With this scale, every agent's evidence enters with a coefficient between 0 and 1. Because each $\Lambda_j$ is positive semidefinite, $P_i^{(t)}$ never exceeds the centralized precision, so the pooled error bars are conservative at every round. As mixing completes, the coefficients approach 1 and the pooled posterior equals the centralized sparse Gaussian process.

This rule removes the product-of-experts pathology by construction. An agent with no data near location $x$ contributes zero evidence there, rather than a copy of the prior, so no heuristic weight is needed to silence it. Streaming data is handled by grouping evidence into time buckets by age, with each bucket scaled by its own $s$. Buckets older than the mixing time are merged into the fully scaled pool. Memory is therefore a few $m\times m$ matrices, independent of mission length or graph diameter.

Relation to Prior Work

Several papers motivate this proposal by identifying the gap. Distributed learning with infinitely many hypotheses extends geometric pooling to continuous spaces as mirror descent but states that transmitting continuous densities is impractical. A shared sparse basis is the practical parameterization it leaves open. Non-Bayesian Social Learning With Uncertain Models shows that geometric pooling converges to the $N$-th root of the centralized evidence. This is the reason a rescaling is required, and our contribution is choosing that rescaling so the transient phase is safe. Network independent rates in distributed learning separates a phase governed by the network's topology from an asymptotic phase that does not depend on topology. Our $s_t$ targets exactly the first phase, which those results treat as a cost and not as a calibration hazard.

Three papers supply methods we reuse and also define what we must beat. Generalized Product of Experts multiplies Gaussian process posteriors with entropy-based weights at a central fusion node. Generalized Robust Bayesian Committee Machine proves that plain products of experts are inconsistent and repairs them with a global expert connected in a star topology. Healing Products of Gaussian Process Experts shows that aggregation must happen in the space of the latent function, not in the space of noisy observations. All three pool posteriors through a coordinator. We pool evidence over multi-hop graphs, and their consistency results then follow from additivity, with no global expert required.

The closest robotics work is Dense Incremental Metric-Semantic Mapping. It already exchanges sparse Gaussian process information matrices and vectors. It avoids double counting ("echoes") by tracking update histories, at a memory cost that grows with history and graph diameter. We replace exact tracking with a provable conservative bound, so the question becomes whether the bound costs little enough accuracy.

Two other papers constrain the design rather than supply it. Cooperative Control of Uncertain Multiagent Systems via Distributed Gaussian Processes averages predictions arithmetically, weighting by variance, and caps local data at 100 points. Distributed multi-agent target search and tracking runs linear consensus on truncated basis coefficients and feeds the resulting maps to a policy learned by reinforcement learning. Both show that sharing beats isolated learning. Neither guarantees calibrated variances during consensus, and that guarantee is what exploration under partial observability depends on.

Research Program

The work has three parts.

The theory part proves the conservative-precision property for streaming bucketed evidence on static graphs. It then extends the proof to time-varying graphs, replacing $\lambda^t$ with a product of per-round contraction bounds. It also bounds the error of the pooled mean, which is not conservative in general.

The systems part builds the protocol with a shared inducing basis fixed in advance. This is either a grid over the known operating region or random Fourier features generated from a common seed. Kernel hyperparameters are shared and fit on a pilot dataset; heterogeneous hyperparameters are left explicitly out of scope. Agents estimate $N$ and a bound on $\lambda$ through one-time gossip.

The closed-loop part places the protocol inside a simple, deliberately unlearned exploration rule: each agent moves greedily toward the location of largest pooled posterior variance. Any change in mission performance can then be attributed to the beliefs and not to policy learning. This tests the core partial-observability claim. Overconfident pooling should make robots skip regions they wrongly think are known. Under-scaled pooling should make robots revisit regions redundantly.

Baselines

The naive reference is isolated local sparse Gaussian processes with no communication, which the Lederer et al. paper above already shows to be worse than sharing.

The accuracy ceiling is a centralized sparse variational Gaussian process using the same inducing points and all data. On UK apartment prices, Generalized Product of Experts reports root mean square error 0.456 for tree-based gPoE and 0.426 for stochastic variational inference. Our pooled model after mixing should reach about 0.426 under K-means partitions. Under trajectory-induced partitions, it should clearly beat gPoE run with one-hop neighbour pooling.

The strongest decentralized competitor is the echoless protocol of Zobeidi et al. It is exact, so beating it means matching its error within statistical noise while using less memory and fewer bytes per agent. We will measure where that crossover happens as the graph diameter grows. No reported scalar result was found on Lacuna for the echoless protocol, so we will reimplement it.

The decentralized versions of gPoE and barGP pool posteriors among neighbours using their published weightings, with the same inducing budget, communication rounds and hyperparameter tuning budget. On Parkinsons, Healing Products of Gaussian Process Experts reports negative log predictive density 0.10 for softmax-variance weighting versus about 1.05 for uniform products. These figures give a scale for the calibration gap we expect posterior pooling to show once partitions are spatially contiguous.

The arithmetic variance-weighted rule of Lederer et al. is included as the linear-pooling control.

The decisive ablations isolate the scale factor. We compare $s_t=N$, the consensus-information-filter choice, which is exact asymptotically but double counts early; $s_t=1$, which is conservative like covariance intersection but underconfident by a factor of $N$ forever; and our mixing-aware $s_t$.

Why not just…?

Why not just use the exact echoless protocol? It is the real threat on small, well-connected teams, where tracking histories is cheap and exactness is free. Our bet is that its memory grows with diameter and mission length while ours does not. The scaling experiment on line and grid graphs with 4 to 64 agents settles this. If the echoless protocol's overhead stays small at realistic team sizes, we will report that our method's value is limited to large or sparse networks.

Why not just multiply neighbours' posteriors with gPoE or barGP weights? Those weights were designed to repair a centralized product of posteriors. Over many hops, each agent's posterior already contains the prior and possibly echoes of its neighbours' evidence, so the weights must compensate for double counting they were never designed for. Trajectory-induced partitions on UK apartment prices and on the Clarence Centre computational fluid dynamics dataset (indoor carbon dioxide) are the test. We predict a negative log predictive density gap like the uniform-versus-weighted gap reported in Healing Products of Gaussian Process Experts, now running against posterior pooling.

Why not just flood raw observations so every agent fits the full model? When per-agent data rates are low, this is simpler and exact. We will report bytes sent against error. Evidence messages cost $O(m^2)$ per round no matter how much data an agent holds, so they should win once each agent's data exceeds roughly $m^2$ values. We will locate that crossover rather than assume it.

Why not just scale by $N$ and accept a brief transient? In exploration, the transient is exactly when decisions are made. The closed-loop experiments measure how often agents skip truly unvisited regions under $s_t=N$ compared with our scaling.

Proposed Experiments

The static experiments use kin40k, SARCOS and UK apartment prices. Each dataset is partitioned two ways: by K-means, for comparability with published aggregators, and by simulated robot paths over the input space, which is the partial-observability regime. We vary graph topology and the number of gossip rounds.

The metrics are standardized mean squared error, root mean square error, and negative log predictive density. We also report the fraction of test points whose pooled variance falls below the centralized variance; our theory says this fraction should be zero.

The closed-loop experiments use the modified multi-agent particle environment from Kim et al. and a field derived from the Clarence Centre computational fluid dynamics data. The measures are field mean squared error over time, time to locate targets, and redundant coverage.

The decisive check has two parts. On trajectory partitions, mixing-aware evidence pooling must never undercut centralized variance, must match the echoless protocol's error with memory that does not grow, and must beat neighbour-pooled gPoE in log predictive density. If the scale $s_t=N$ produces equal exploration performance despite its transient overconfidence, the conservative scaling is unnecessary in practice, and we will say so.