Efficient and reliable IMRT treatment planning is challenging even when using only a single frozen-in-time CT scan of anatomic structures. The challenge is intensified in 4-D treatment planning, which is based on highly expanded imaging datasets that provide views of structure shape and position shifts over time. Incorporating these expanded datasets into the treatment planning process has the potential to yield better treatment plans, but at the same time results in models and optimization problems that are several magnitudes larger than those associated with traditional single-time-period planning. And along with the increase in problem size, there are additional sources of uncertainty and error (e.g., uncertainties in breathing trajectories, errors in organ contour outlining related to the increased number of images). Treatment planning methods must therefore be developed that can accommodate the increased problem size, and at the same time compensate for the errors and uncertainties. In this talk, we describe various mathematical models for such robust and adaptive large-scale planning methods. Their mathematical complexity will be analyzed, and theoretical results will be described. We will demonstrate that our methods allow a significant reduction in mean dose to normal tissue, and in some cases, higher dose to tumor volume.