What was your approach to benchmarking an adversarial agent? This is an open problem that I came across (in a different domain), as the search space can be rea

What was your approach to benchmarking an adversarial agent? This is an open problem that I came across (in a different domain), as the search space can be really wide. It's hard to measure results for non-trivial tasks. Would be really interested if you can share your eval approach :)