A new estimator lets researchers compute stable strategies for cybersecurity simulations without running every costly test scenario.
The technique, called Regret-Weighted Payoff Sampling, targets a bottleneck in policy-space response oracles (PSRO), a method for finding strong attacker and defender strategies through repeated simulation. Normally, every entry in the resulting payoff matrix - essentially a scorecard of how each strategy fares against every other - requires slow Monte Carlo rollouts, meaning thousands of simulated skirmishes, to estimate. RWPS instead simulates only the matchups that actually shift the equilibrium and fills in the rest with a surrogate model trained on earlier results. Tested on three 21x21 games, including an asymmetric Colonel Blotto scenario, it simulated as little as 18% of the matrix for simpler games and up to 82% for the hardest one, while beating three rival sampling strategies at finding low-exploitability strategies on a matched compute budget.
The real win is knowing in advance which matchups are worth simulating and which aren't. Cybersecurity simulators like CyGym and ANSG are expensive to run at any meaningful scale, and most game-theoretic security research has been budget-constrained rather than method-constrained. A sampling approach with a provable error bound could let researchers model bigger, more realistic attacker-defender scenarios without a proportional jump in compute costs.
Whether that translates into better real-world defenses is a separate question. This is a paper about making the math cheaper, not about a new defense capability - the underlying simulators still have to reflect real attacker behavior for any of it to matter outside a benchmark.