Tumor-immunotherapy dosing drops the forward-backward sweep: 14 of 16 scenarios stalled, backprop training converged on all
Synopsis
Optimal control of tumor-immune dynamics is usually solved by a forward-backward sweep on the Pontryagin necessary conditions, but that iteration can cycle between two saturation patterns instead of converging when the control saturates. Representing the control as a neural network and the cost functional as the training loss makes every stationary point of the loss satisfy the continuous first-order necessary conditions to a discretisation-controlled accuracy. On a 4×4 grid of scenarios from a tumor-immune model, the standard sweep failed on 14 of 16 scenarios, while network training reached a stable optimum on every scenario with a final cost 2 to 10 times lower than the converged damped sweep.
Interpretation
Optimal dosing schedules for tumor immunotherapy no longer have to come from a forward-backward sweep iteration; the control function is a neural network trained directly on the cost functional. The prior standard iterates the Pontryagin maximum principle necessary conditions by a forward-backward sweep, an iteration with known instability on stiff systems, near bifurcations, and where the optimum pushes the control onto a constraint boundary. On a 4×4 grid of scenarios drawn from a published immune-cancer dynamics model, the nominal sweep (damping 0.5, iteration cap 30) failed on 14 of 16 scenarios, judged by a relative-change metric staying above 0.05 through the last 8 iterations; the two converging scenarios were both high-tumor initial conditions at low γ.
The trained controls converged stably on every scenario and reached a lower cost. Lowering damping to 0.1 with adaptive halving made the sweep converge on 15 of 16 scenarios, but raised the iteration cap from 30 to 500, took about 87 minutes in the worst case, and still left one scenario beyond a two-hour budget. On the same grid, network training met the convergence criterion on all 16 scenarios in 4 to 32 minutes each; on the 15 scenarios where the damped sweep converged, the trained final cost was 2 to 5 times lower, and more than 10 times lower in several high-effector and high-γ scenarios.
Changing the cost functional requires changing only the loss, not the training pipeline. Under the canonical cost, the trained schedule is a roughly 10-day pulse followed by withdrawal after tumor clearance, because the cost has no reward for a sustained effector population. At the delayed-response scenario with γ=37.417 and baseline initial conditions, an induction-maintenance recurrence penalty, a hard cumulative-dose budget, and a chemotherapy-plus-immunotherapy variant with a third control all reused the same 200-step Adam plus 30-step L-BFGS pipeline, taking 7 to 25 minutes.
Perspective
The demonstration covers one published immune-cancer dynamics model, comparing the sweep and network training on 16 scenarios formed from 4 bifurcation-parameter values and 4 initial conditions, plus three cost-functional variants at the delayed-response scenario. The intended reader is a modeling researcher comfortable with optimal control and numerical ODE solving. The authors list HIV treatment, vaccination control on epidemic models and ecological harvesting as possible extensions and state that these were not validated here.
Training reaches a local stationary point of the loss, not a global optimum; the authors suggest multi-start seeds or a Hamilton-Jacobi-Bellman solver to obtain a global cost for comparison. The toxicity term is a quadratic penalty and does not capture threshold-style adverse events at high checkpoint-inhibitor doses. Cost weights are scale-balanced rather than calibrated to clinical quality-adjusted life year data. The comparison is against the forward-backward sweep only, not direct collocation or indirect semi-discretisation.