TR2024-167

TR-BEACON: Shedding Light on Efficient Behavior Discovery in High-Dimensional Spaces with Bayesian Novelty Search over Trust Regions

    •  Tang, W.-T., Chakrabarty, A., Paulson, J.A., "TR-BEACON: Shedding Light on Efficient Behavior Discovery in High-Dimensional Spaces with Bayesian Novelty Search over Trust Regions", Advances in Neural Information Processing Systems (NeurIPS), December 2024.
      BibTeX TR2024-167 PDF
      • @inproceedings{Tang2024dec,
      • author = {Tang, Wei-Ting and Chakrabarty, Ankush and Paulson, Joel A.}},
      • title = {TR-BEACON: Shedding Light on Efficient Behavior Discovery in High-Dimensional Spaces with Bayesian Novelty Search over Trust Regions},
      • booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
      • year = 2024,
      • month = dec,
      • url = {https://www.merl.com/publications/TR2024-167}
      • }
  • MERL Contact:
  • Research Areas:

    Artificial Intelligence, Control, Machine Learning, Multi-Physical Modeling, Optimization

Abstract:

Novelty search (NS) algorithms automatically discover diverse system behaviors through simulations or experiments, often treating the system as a black box due to unknown input-output relationships. Previously, we introduced BEACON, a sample-efficient NS algorithm that uses probabilistic surrogate models to select inputs likely to produce novel behaviors. In this paper, we present TR-BEACON, a high-dimensional extension of BEACON that mitigates the curse of dimensionality by constructing local probabilistic models over a trust region whose geometry is adapted as information is gathered. Through numerical experiments, we demon- strate that TR-BEACON significantly outperforms state-of-the-art NS methods on high-dimensional problems, including a challenging robot maze navigation task.