Models (esms): a Multi-task Learning Perspective

Current evaluations of GCMs, as detailed by the IPCC AR5, show that different GCMs often illustrate different skills for different tasks. The IPCC uses an average or multi-model ensemble (MME) for future projections. In this work, we consider evaluating and combining GCMs for multiple tasks from a multi-task learning (MTL) perspective. MTL is an approach to machine learning that learns multiple problems simultaneously thus leading to a better model by taking advantage of the commonality among the tasks [1], [2]. But unlike traditional MTL, for our problem the task relationships are not known initially. Further, each GCM has multiple runs corresponding to different initial conditions which need to be suitably considered. In this work, we present approaches to multitask sparse structure learning (MSSL) which estimate task relationships along with learning suitable multi-model combinations for each task, and explore ideas for handling multiple initial condition runs for GCMs.