Hyperbolic Inverse Reinforcement Learning
Hyperbolic distance parameterization of rewards boosts IRL accuracy in low-data regimes and reveals non-separable conjunctive structures.
The authors introduce Hyperbolic IRL, which models rewards using hyperbolic distance between feature vectors and learned parameters alongside Euclidean distance. Tested against linear and nonlinear baselines in deterministic and stochastic settings, the method matches state-of-the-art performance while excelling in low-data and dense environments. Applied to human museum navigation, it uncovers reward dependence on both individual distances and non-separable conjunctions.
- Comparable performance to SOTA on benchmarks
- Superior in low-data limit and dense environments
- Reveals non-separable conjunctions in rewards
- Improves predictive accuracy and interpretability
- Validates on human navigation task
From the original article
1 min read at arxiv.orgInverse reinforcement learning (IRL) seeks to infer the reward function underlying observed behavior, but its predictive accuracy and interpretability depend on how that reward is computed from the underlying features.
By Iulia Rusu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.