MAVRL: Learning Reward Functions from Multiple Feedback Types with Amortized Variational Inference
This addresses the problem of reward learning from multiple feedback types for reinforcement learning practitioners, offering a scalable method that eliminates manual tuning and improves robustness, though it is incremental in combining existing feedback types with a novel inference framework.
The paper tackles the challenge of jointly learning reward functions from heterogeneous feedback types like demonstrations and comparisons by formulating it as Bayesian inference over a shared latent reward function, using an amortized variational inference approach that avoids manual loss balancing. The result shows that jointly inferred reward posteriors outperform single-type baselines, exploit complementary information, and yield more robust policies across benchmarks.
Reward learning typically relies on a single feedback type or combines multiple feedback types using manually weighted loss terms. Currently, it remains unclear how to jointly learn reward functions from heterogeneous feedback types such as demonstrations, comparisons, ratings, and stops that provide qualitatively different signals. We address this challenge by formulating reward learning from multiple feedback types as Bayesian inference over a shared latent reward function, where each feedback type contributes information through an explicit likelihood. We introduce a scalable amortized variational inference approach that learns a shared reward encoder and feedback-specific likelihood decoders and is trained by optimizing a single evidence lower bound. Our approach avoids reducing feedback to a common intermediate representation and eliminates the need for manual loss balancing. Across discrete and continuous-control benchmarks, we show that jointly inferred reward posteriors outperform single-type baselines, exploit complementary information across feedback types, and yield policies that are more robust to environment perturbations. The inferred reward uncertainty further provides interpretable signals for analyzing model confidence and consistency across feedback types.