A receiver that learns when and where to listen.
Radars show up for milliseconds at a time, and a receiver can only tune to one slice of spectrum at once. Eklavya learns each radar's rhythm from what it hears, predicts the next flash, and spends its listening time where it will catch the most.
Why scanning is hard
Radars flash briefly
A rotating radar points at you for a few milliseconds every few seconds. Each flash is the only chance to hear it.
The receiver hears one slice at a time
The band is wide but the receiver tunes to a narrow slice. Every look is a bet on where the next flash will be.
A fixed sweep gets locked out
A sweep repeats on a timer. A radar whose rhythm lines up with it is always flashing somewhere else, and is never seen.
Six receivers, same radars
Every mission runs all of them against identical ground truth, so the only difference is how each one decides where to listen.
Eklavya
Learns each radar's rhythm, predicts its next flash, and spends listening time where it pays off most.
Double-DQN
A deep reinforcement learning agent that learns where to listen purely by trial and error.
Greedy bandit
Keeps going back to the frequencies that paid off before. Simple, but it stops exploring.
Raster sweep
The traditional fixed sweep across the band. Predictable, so radars in step with it are never seen.
Random
Picks a random slice of spectrum every look. The floor any scheduler has to beat.
Oracle
Cheats: it sees the true radar schedule. Shows the best score that is physically possible.
See it on a real mission
Pick a scenario, run every receiver at once, and watch the flashes they catch add up.