Publications & Research
Research on adaptive AI safety, safe continual reinforcement learning, and safety under nonstationarity.
Publishing professionally as Tim Tomashevskiy; some earlier publication records appear under Timofey Tomashevskiy or Timothy Tomashevskiy.
Peer-Reviewed / Accepted
Conference publications
Proactive Context-Forecasted Safety Constraints for Nonstationary Reinforcement Learning
Introduces proactive safety-constraint generation from inferred and forecast environmental context, allowing an agent to adapt constraints before changing conditions lead to unsafe behavior.
A Safety Architecture for Centralized E/E Architectures
Earlier work on architectural safety for centralized automotive electrical/electronic systems, forming part of the functional-safety foundation for my current research on learning-enabled safety.
Preprints
Current research
Safe Continual Reinforcement Learning under Nonstationarity via Adaptive Safety Constraints
Presents LILAC+, a predictive safety layer combining context-based, adaptation-speed, and soft-to-hard safety constraints for policy-level adaptation and runtime enforcement.
From Cumulative Constraints to Adaptive Runtime Safety Control for Nonstationary Reinforcement Learning
Develops CPSS, a runtime mechanism that projects cumulative safety requirements into local action-level constraints, connecting longer-horizon safety budgets to enforceable runtime control.
Adjustment Speed as a Safety Constraint for Nonstationary Reinforcement Learning
Frames adaptation feasibility as a safety problem by comparing the rate of environmental change with an agent's effective adaptation capacity and proactively tightening behavior when safe adaptation may be too slow.
Safe Continual Reinforcement Learning Methods for Nonstationary Environments. Towards a Survey of the State of the Art
Surveys safe continual online reinforcement learning under nonstationarity, organizing methods around safe learning mechanisms, adaptation to distribution shift, and safety-constraint formulation.
Work in Progress
Ongoing research
SafeShiftRL: Benchmarking Safety in Continual Reinforcement Learning under Nonstationarity
Developing a modular benchmark methodology for evaluating safety, nonstationarity, and adaptive constraint validity across stationary, seen-nonstationary, and unseen-nonstationary regimes.