Conceptmodel exploitation
Model Exploitation
A failure mode in model-based reinforcement learning in which the planner exploits prediction loopholes in the world model to seek states that are infeasible in reality but predicted to yield high reward. It is hard to detect within the training distribution and fails rapidly under distribution shift.
Timeline (last 90 days)
Oct 12
The cost of freezing the world model training strategy is: if the world model itself has bias, the policy will overfit to that bias, i.e., the phenomenon of 'model exploitation' occurs.
Unverified50%