Isolate the effect of reaction delay and decision frequency

Determine how much reaction delay and decision frequency contribute to the reported performance differences between Faynt and benchmark policies with delayed action execution.

Background

Faynt uses zero added policy delay, whereas several Slippi-AI comparison checkpoints queue actions for 18, 21, or 24 game frames; the HAL comparison also uses a different decision frequency. The benchmark protocol therefore compares policies with different timing characteristics.

The paper reports strong performance differences but does not experimentally separate the effects of reaction delay from those of decision frequency. Establishing these effects would require controlled evaluations in which the policies operate under matched and systematically varied timing conditions.