Dead rats, dopamine, performance metrics, and peacock tails: Proxy failure is an inherent risk in goal-oriented systems | Behavioral and Brain Sciences | Cambridge Core

There's work on this, of course e.g. this by @dryohanjohn.bsky.social et al is v useful. But equally e.g. the reward function is often under scrutinized in robotics RL, a hiding place for hand definition of baroque constraints wangling brittle solutions www.cambridge.org/core/journal... Dead rats, dopamine, performance metrics, and peacock tails: Proxy failure is an inherent risk in goal-oriented