Validation hides the next domain
Policies can look reliable in familiar simulation while small changes in dynamics or observation cause abrupt performance loss.
Deep reinforcement learning / robustness
Domain-shift quantification.
A dissertation project about measuring whether a policy remains suitable for an upcoming environment before deployment reveals the failure.
Interactive experiment map
Observe the policy across controlled shifts.
Separate changes to dynamics from changes to observation, collect behaviour, and let the meta-model estimate suitability.
Case study
Robustness as a prediction problem
Learn from controlled shifts
Generate shifted environments, collect policy behaviour and returns, then fit a second-order model linking shift signals to policy suitability.
Keep experiment and argument together
The public repository contains the DomainShift experiment code, tests, paper material, and the associated dissertation.
Experiment detail
A four-stage robustness probe
Train a source policy
Establish behaviour in the environment the agent sees during learning.
Parameterize shifts
Alter dynamics and observations separately so the source of change remains legible.
Collect rollouts
Measure trajectories, state visitation, and returns under each controlled change.
Estimate suitability
Use a meta-model to predict whether the policy and upcoming domain are compatible.
Evidence boundary
A research framework for reasoning about shift.
The project motivates and implements intrinsic shift quantification. It should be read as dissertation research rather than a universal guarantee of out-of-domain policy safety.