Arguably one might need to capture a bit more data describing the context and outcome associated with each recorded "deployment" of each deck. With only a coup

Arguably one might need to capture a bit more data describing the context and outcome associated with each recorded "deployment" of each deck. With only a couple of hundred samples, and a highly complex process with many poorly understood interacting factors, and loads of missing information, throwing neural networks at it sounds like a fun exercise in overfitting! Could probably get further by getting hold of a few experts who have strong prior beliefs about what factors may be relevant, and then carefully running some statistical analysis / additional experiments! "statistics" and domain expertise - not trendily branded, probably still a great idea when you merely have a few hundred samples in a crazily ill-defined high dimensional space.