The Battery That Wore Itself Out: Classical Reinforcement Learning

Deep End Era. Engineering with Sankalp · Oct 28, 2026 · 52m

A fictional grid battery's new RL trading agent beats the human-written rules by 18 percent in simulation by cycling the battery dozens of times a day, which would wear it out in four years instead of fifteen. Trading lead Olu Adebayo and ML engineer Hanna Lindqvist work out why the agent did exactly what it was told. Part fourteen of The Deep End's Machine Learning series covers classical RL from Bellman to AlphaGo, and the fixes for reward hacking: honest costs in the reward, hard limits outside the agent, honest simulators and shadow mode. Not LLM post-training (ep116). Cross-references…