Stas Tiomkin

About

Stas Tiomkin is a researcher advancing the frontiers of deep reinforcement learning (RL), with a focus on building agents that are robust, efficient, and secure. His work tackles fundamental challenges in making RL deployable in the real world. Tiomkin’s most influential contribution, with 17 citations, is his 2020 paper "Dynamics Generalization via Information Bottleneck in Deep Reinforcement Learning," which addresses the critical problem of overfitting in RL agents, preventing them from adapting to unseen environments. He further explores the underexamined area of policy privacy in "Preventing Imitation Learning with Adversarial Policy Ensembles," proposing methods to protect proprietary policies from being cloned by external observers. To overcome the persistent challenge of sparse rewards in long-horizon tasks, Tiomkin introduced an effective reward-shaping method using predictive coding in his 2019 work. His more recent research on multi-objective policy gradients with topological constraints (2022) offers a novel framework for encoding ordered preferences and safety constraints. Through these contributions, Tiomkin is shaping a future where RL agents are not only more capable but also safer and more private.

Research Focus

Key Achievements

Top Papers

1. 1
   Dynamics Generalization via Information Bottleneck in Deep Reinforcement Learning

   17 citations · 2020
2. 2
   Preventing Imitation Learning with Adversarial Policy Ensembles

   4 citations · 2020
3. 3
   Predictive Coding for Boosting Deep Reinforcement Learning with Sparse Rewards

   4 citations · 2019
4. 4
   Multi-Objective Policy Gradients with Topological Constraints

   2 citations · 2022

Key Collaborators

Contact & Links