quantdata.dysanos.grid.rndysanos.run_no_offset_comparison#
Compare no-offset path-OCE optimizers on one shared simulated universe.
The comparison deliberately separates two questions:
three neural strategies use a constant-rate AdamW optimizer and transaction costs; and
one deterministic-time power-exponential strategy is a cost-free convex benchmark solved by SciPy L-BFGS-B over a GPU Torch objective and gradient.
All four workers receive the same read-only universe through shared memory and
are assigned independent CUDA devices by quantdata.torch.DeviceLock.