Adversarial Graph Traversal

Suppose a Bayesian agent seeks to traverse a graph. Each time she crosses an edge, she pays a price. The first time she reaches a node, there is a payoff. She has an opponent who can reduce the payoffs. This paper uses adversarial risk analysis to find a solution to her route selection problem. It shows how the traveler is advantaged by having an accurate subjective distribution over the costs/payoffs and by having a Bayesian prior for her opponent’s strategic choices. The results are relevant to military convoy routing, corporate competition, and certain games.

💡 Research Summary

The paper introduces a novel framework for sequential graph traversal under uncertainty and adversarial interference. A Bayesian traveler moves through a graph, incurring a stochastic cost each time an edge is crossed and receiving a stochastic payoff upon first visiting a node. The traveler’s prior beliefs about the joint distribution of costs and payoffs are modeled with a Normal‑Inverse‑Wishart (NIW) distribution, which remains closed under conditioning; integrating out the precision matrix yields a multivariate t (MV‑t) distribution that provides tractable expectations for unobserved components.

An opponent can, once per turn, reduce the payoff at one neighboring node by a known fraction (δ = 0.3). Two opponent types are considered: Type 0 (greedy) reduces the highest‑value payoff, while Type 1 (strategic) reduces the second‑highest, anticipating the traveler’s expectation of a greedy reduction. This corresponds to level‑k thinking with k = 0 or 1. The traveler holds a Beta(α, β) prior on the probability π that the opponent is Type 0, and updates this prior using Bayesian evidence derived from observed payoffs.

For each neighboring node i, the traveler computes q₁ᵢ (probability that i is the best choice) and q₂ᵢ (probability that i is the second‑best) using the MV‑t predictive distribution. Observed payoff x is compared against two conditional densities: f(x | reduced) and f(x | not reduced). The posterior probability that the payoff was reduced, γ = P(reduced | x), is then used to update the Beta prior to Beta(α + γ, β + 1 − γ).

Because the mixture of NIW components doubles with each new observation, the authors propose pruning low‑weight components to keep computation tractable. They also introduce an “uncertainty policy” that adapts the search horizon based on the relative change τ in conditional means after new data: τ > 20 % limits search to immediate neighbors, 10 % < τ ≤ 20 % expands to two‑step paths, 5 % < τ ≤ 10 % to three‑step paths, and τ ≤ 5 % to four‑step paths. This dynamic depth control mitigates the NP‑hard nature of the underlying stochastic shortest‑path problem.

The methodology is evaluated on a 6 × 6 grid graph, comparing three scenarios: (1) no opponent, (2) opponent present but ignored, and (3) opponent modeled with Bayesian updating. Results show that incorporating the adversary’s behavior substantially increases cumulative net reward, especially when the traveler’s prior on opponent type is accurate or only mildly misspecified. The adaptive strategy outperforms a fixed‑path approach, and the pruning technique keeps runtime feasible.

Overall, the study demonstrates that combining adversarial risk analysis, Bayesian hierarchical modeling, and level‑k reasoning yields a powerful decision‑making tool for applications such as military convoy routing, competitive corporate planning, and security‑oriented games, where agents must operate under uncertainty while anticipating strategic interference.