Description
Reinforcement learning trains a policy that takes actions in an environment to maximise cumulative reward. In the WiFi-sensing context it shows up mostly in two niches: adaptive channel / antenna selection for sensing, and crowd-control / evacuation policies on top of agent-based-model simulators. It is not the dominant CSI-sensing learning paradigm.
When it's used
- Adaptive sensing schedules (when to probe, which channel)
- Crowd-management policy learning over simulator rollouts
- Resource allocation in ISAC waveform design
Limitations
- Sample inefficiency unless paired with simulators
- Reward design is the hard part
- Sim-to-real gap for CSI rewards