You can't. That is, the optimal is just the optimal value at each point in your sample. You cannot do better than that, at least, not unless you are willing to build a model of the process.
So, what are you willing to do in terms of a model? What do you know? What are you willing to assume? If you cannot build a model, then you cannot do any form of optimization that will have some chance of being better than the best point from your data.
If your data is at all noisy, then don't expect to do well of course. But one option might be to use an interpolating spline surface model. Again, noise will totally destroy any spline fit, making it useless for this purpose.