Conformal Prediction for Uncertainty-Aware Planning with Diffusion Dynamics Model

@inproceedings{sun_conformal_2023,
  address = {New Orleans, U.S.},
  title = {Conformal {Prediction} for {Uncertainty}-{Aware} {Planning} with {Diffusion} {Dynamics} {Model}},
  abstract = {Robotic applications often involve working in environments that are uncertain, dynamic, and partially observable. Recently, diffusion models have been proposed for learning dynamics models, demonstrating a strong ability to overcome the multi-modal action distribution, high-dimensional output space, and training instability challenges. The uncertainty of such dynamics models is thus a critical factor to consider when using them for planning. In this paper, we quantify the uncertainty of dynamics models using Conformal Prediction (CP), which is an effective technique for constructing prediction sets that achieve valid coverage. With its distribution-free nature and statistical guarantee from finite samples, CP ensures that if the training and testing data are exchangeable, then the target lies within a prediction set for any predefined target level. When learning to plan under uncertainty, we connect conformal prediction to planning and propose PlanCP to optimize the model by minimizing the uncertainty. Additionally, we assess the effectiveness of uncertainty sets by comparing their coverage and optimization performance. Furthermore, during the test, PlanCP can also measure the model uncertainty. We evaluate our algorithm on various planning tasks and model-based offline reinforcement learning tasks and show that it reduces the uncertainty of the learned dynamics model. As a by-product, our algorithm PlanCP outperforms prior algorithms on existing offline RL benchmarks and two challenging continuous planning tasks. Our method is highly flexible and can combine with most model-based planning approaches and produces uncertainty estimates of the dynamics model.},
  language = {en},
  urldate = {2023-12-10},
  booktitle = {Thirty-seventh {Conference} on {Neural} {Information} {Processing} {Systems} ({NeurIPS} 2023)},
  publisher = {NeurIPS},
  author = {Sun, Jiankai and Jiang, Yiqi and Qiu, Jianing and Nobel, Parth Talpur and Kochenderfer, Mykel and Schwager, Mac},
  month = dec,
  year = {2023},
  note = {Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023)},
  keywords = {conformal prediction, decision making},
  month_numeric = {12}
}