Two-learner problem

The two-learner problem names the fundamental difficulty of co-adaptive BCI: two adaptive agents — the machine decoder and the biological user — update their parameters at the same time, each optimizing against a target that the other is moving. Formally it is a coupled, non-stationary, and generally non-convex optimization; the joint system can converge to good performance, settle into a mediocre equilibrium, oscillate, or diverge, depending on how the two learning processes are scheduled relative to each other.

Two intertwined issues make it hard. First is credit assignment: when performance shifts, it is often impossible to attribute the change to the decoder update or to a change in the user's neural strategy, which frustrates both diagnosis and principled control of adaptation. Second is stability: if the decoder adapts as fast as, or faster than, the user, the two can chase each other, and an adaptation that improves an instantaneous objective can destabilize the loop. Most practical designs resolve this by deliberately breaking the symmetry — letting one learner move much more slowly, or freezing it — so the other has a stable target.