Modeling Blackbox Agent Behaviour via Knowledge Compilation. Muise, C.; Wollenstein-Betech, S.; Booth, S.; Shah, J.; and Khazaeni, Y. In The AAAI 2020 Workshop on Plan, Activity, and Intent Recognition, 2020.
abstract   bibtex   
Understanding how agents behave in an environment is a corner-stone for interpretable AI. In this work, we focus on capturing the policy an agent is following without placing any assumptions on how that policy is actually implemented. From a corpus of state-action pairs, our task is to build a compact and diagnosable representation of the mapping from states to actions. We appeal to modern knowledge compilation techniques for this task and demonstrate empirically how this approach outperforms the previous state of the art. We further create an interactive inference on the compiled representation to get an intuitive sense of the policy.
@inproceedings{muise-pair20,
  title={Modeling Blackbox Agent Behaviour via Knowledge Compilation},
  author={Christian Muise and Salomón Wollenstein-Betech and Serena Booth and Julie Shah and Yasaman Khazaeni},
  booktitle={The AAAI 2020 Workshop on Plan, Activity, and Intent Recognition},
  year={2020},
  abstract={Understanding how agents behave in an environment is a corner-stone for interpretable AI. In this work, we focus on capturing the policy an agent is following without placing any assumptions on how that policy is actually implemented. From a corpus of state-action pairs, our task is to build a compact and diagnosable representation of the mapping from states to actions. We appeal to modern knowledge compilation techniques for this task and demonstrate empirically how this approach outperforms the previous state of the art. We further create an interactive inference on the compiled representation to get an intuitive sense of the policy.}
}
Downloads: 0