Sensor networks stop guessing only locally: nodes first find same-state neighbors, then recover the whole network

Each agent in a sensing network can now infer the hidden states of all agents in the network, not only its own. Earlier decentralized inference either assumed a single state shared across the network or let each agent estimate only its own state, using cooperation mainly to improve that local task. In a synthetic network of 20 agents with 5 Gaussian means, after 50 cluster-discovery iterations the label vectors stabilized to a partition matching the connected same-state components, for example all agents labeling 1, 6, 11, and 16 as one cluster.

The method runs in two phases: discover clusters first, then perform social learning at the cluster level. An earlier multivariate-belief formulation for network-wide states required each agent to maintain and exchange beliefs over all possible network-wide state configurations, whose dimension grows exponentially with network size. In cluster discovery, each agent performs a local Bayesian update of its own-state belief, compares beliefs with neighbors to merge same-state neighbors under a threshold epsilon, and joins neighbors' intermediate partitions; in the second phase each agent keeps one probability vector per discovered cluster, updating its own cluster from local observations and other clusters by geometric pooling of neighbors' intermediate beliefs.

The information exchanged scales with the number of clusters rather than the number of agents or network-wide configurations. Maintaining and exchanging decision statistics per agent would make communication cost grow with network size. The synthetic experiment with 20 agents converged to 5 reference clusters, so each agent maintained 5 cluster beliefs in the second phase; the real gas-sensor experiment with 10 agents and 4 gases converged to 4 clusters.

On real gas-sensor data, agents recovered the network-wide gas-state configuration using only local measurements and neighbor communication. The synthetic experiment used a known Gaussian observation model, whereas the real data has no closed-form likelihood and required likelihood ratios built from local classifier outputs. The experiment used the UCI Twin Gas Sensor Arrays dataset with 10 agents, each with a four-channel chemical sensor and hidden states among CO, ethanol, ethylene, and methane; cluster discovery used the first 5 feature vectors with threshold epsilon of 10 to the minus 3, and by time step 20 each agent's belief in every target agent's true gas state was close to one.

The method is validated under fixed hidden states and a fixed communication graph, with agents inferring network-wide states from local observations and neighbor exchanges. The synthetic experiment uses 20 agents and 5 Gaussian means, and the real experiment uses 10 agents and 4 gases, both assuming agents sharing a hidden state form a single connected component in the communication graph. It suits researchers and engineers working on decentralized inference and sensing-network monitoring.

Cluster discovery and network-wide state inference are decoupled, so agents must finish cluster discovery before starting inference. If two agents with different hidden states are incorrectly merged during cluster discovery, the second phase cannot fully correct that error. Cluster discovery depends on jointly choosing the threshold epsilon and the clustering horizon T: a larger threshold merges faster but risks incorrect merges, while a smaller threshold is more conservative but needs a longer phase; increasing T beyond the time needed to stabilize the cluster structure mainly delays the start of the second phase. The current setting targets fixed hidden states and a fixed communication graph.