Submission Types:
Original Work: 4-8 pages in RLC template such as early experimental results.
Published Work: Camera ready version of an accepted conference or journal paper.
NeurIPS work: directly re-submit your NeurIPS work.
Submission information:
All papers accepted to InterpPol can be presented as posters on 9th of August @RLC. There will be NO proceedings of accepted papers at InterpPol. As such, it is possible to send already published work and submitted work to InterpPol.
Reviewing Process:
InterpPol will use this OpenReview page for reviewing. The reviews are single-blind and focused on correctness and fit to topics of interest (see below).
Key dates
Submission deadline: May 23, 2024 AoE
Acceptance notification: May 30, 2024 AoE
Topics of interest (not limited to):
Why learn interpretable policies? What are the interpretability requirements of sequential decision problems,e.g. in the health sector? Are there practical advantages of learning interpretable policies such as decision trees rather than neural networks with respect to RL goal alignment or formal verification? Why looking for intrinsically explainable –interpretable– policies rather than post-hoc explainability?
How to define interpretability? What is an interpretable policy? How to compare different classes of interpretable policies such as trees or programs? How to quantify interpretability of policies with metrics or with (simulated) user studies?
How to learn interpretable policies? What are the advantages of the different learning paradigms (e.g. imitation learning, direct reinforcement learning, or evolutionary methods)? When to use direct interpretable policy learning rather than imitation? Are some interpretable structure easier to learn than others, e.g. trees against programs? What are the connections between RL for combinatorial optimization and interpretable RL?
What sequential problems can be solved with interpretable RL? How can we refine MDPs to be most conducive to learning interpretable policies, especially in terms of interpretable state space representations? How to learn new models over interpretable state spaces? Do all interpretable MDPs have an interpretable solution? How to use reward shaping to trade-off between interpretability and performance of the learned policy?
Formal verification and RL, real-world and applications of RL are also topics of interest.