Robust Structured Prediction for Process Data

Processes involve a series of actions performed to achieve a particular result. Developing prediction models for process data is important for many real problems such as human and animal behavior modeling, psychological evaluation, labor hiring cost assessment, stock investment, human robot interaction and so on. Our contribution presented in this thesis first tractably extends the principle of maximum causal entropy to certain partially observable environments. More specifically, we develop IRL methods for the linear-quadratic-Gaussian system, a well known optimal control problem with partial observability. Furthermore, we investigate process prediction problems under non-stationary settings. One form of this problem is known as covariate shift, where the input distributions for training and testing are different while the mappings from input to output are the same. We propose a robust approach to deal with covariate shift for linear regression as a significant first step to deal with covariate shift for general process prediction tasks. Finally, we introduce a general framework for imitation learning, an important process prediction task where a learner attempts to imitate a demonstrator's behavior from observed demonstration. Our framework enables learning for general evaluation measures and different capabilities between the learner and demonstrator. We demonstrate the effectiveness and show the benefits of our approaches on both synthetic and real datasets.