Research

Our root node questions are:

• Epistemics (E): Developing predictors of agent perceptions with reliable uncertainty estimates.
• Synergetics (S): Developing adaptation algorithms for agents with uncertain perceptions operating in open-world environments.
• Non-asymptotics (N): Predicting lifelong agent performance via mathematical statements.

which synergize via the following research questions:

We reached the following key outcomes:

• (S+N) One can learn an isomorphism of the latent dynamics of a controlled Markov process from hitting time observations, which can be used to train foundation policies. See our IEL algorithm (Boock et al., arXiv, 2026).
• (N) Reconstructing reinforcement learning theory from measure-theoretic foundations explains why deep actor-critics work well. See our MTRL framework (Haussmann et al., arXiv, 2026).
• (E+S+N) Deriving intrinsic reward from well-calibrated uncertainties of the return distributions speeds up adaptation to non-stationary environments. See our DAIF, EPPO, and WSB algorithms.
• (E+S) Evidential uncertainty quantification enhances learning capacity in on-policy continuous control, classification, and image generation. See our EPPO, ETP, and EdVAE algorithms.
• (E+S) PAC-Bayes bounds on parametric return distributions can be used for directed exploration to discover sparse rewards. See our PBAC and PAC4SAC algorithms.
• (E+S) One can exploit the direction of Bellman errors to improve the performance of actor-critic algorithms. See our USAC and AEA algorithms.