What Happens When You Give a Classifier Some Game Sense?

A developer built Signal / Siege, a Rust-based retro defense game, to test how much traditional game AI techniques can improve a classifier's decisions. In a replay-based comparison across 256 seeds with 200ms simulated latency, a simple-rules approach won 112 matches, while adding action scoring and forecasting raised that to 255; in recorded API-backed matches, the Bayesian and hybrid systems survived while the rules-only version lost the base. The developer cautions that the Bayesian reference knows the game's synthetic observation model and that the results do not establish a general ranking of classifiers.

I've been experimenting with a retro defense game in Rust called Signal / Siege. The question behind it is simple: how much can traditional game AI help a classifier? Incoming contacts approach a base. Some are hostile, others should pass safely. The defense can scan, intercept, or activate a shield, but energy is limited. Spending too much on one contact can leave it unprepared for the next. That creates two connected problems: estimating what a contact might be, and deciding what to do about it. The video above shows three approaches: The supporting techniques come from the traditional game AI toolkit. They help the defense weigh consequences, conserve energy, and recognize when waiting could be more useful than acting immediately. Jev's underlying model remains unchanged. In the recorded demonstration, the Bayesian and hybrid systems survived. Jev with simple rules lost the base. We also ran a separate, matched offline comparison across 256 fresh seeds. Using a fixed set of captured Jev responses and 200 milliseconds of simulated latency, the rules approach won 112 matches. Adding action scoring and forecasting brought that to 255. Those were replay-based experiments, not hundreds of live API games. The Jev panels in the video replay actual API-backed matches. The video's three modes have different scanning policies, and the adaptive opponent can respond differently as each match unfolds. The Bayesian reference also knows the game's synthetic observation model, giving it an advantage in this setting. These results don't establish a general ranking of classifiers. What interests me is the division of work: a classifier contributes a judgment, while the surrounding system manages timing, costs, and consequences. Next, I want to test that relationship with richer observations and less predictable scenarios.