Reinforcement learning at MindBubble

In July 2019 I led a reinforcement learning workshop for MindBubble, a metro Atlanta nonprofit that pairs secondary students with free tutoring and STEAM enrichment (i.e., the intellectual stretch a normal school day rarely makes room for). Middle schoolers, laptops, and enough patience for agents before generative AI was anyone's headline.

We taught with OpenAI Gym: an agent acts, the environment returns state and reward, and the loop repeats until policy beats brute force clicks. The lesson I wanted them to keep is the same one I still use when scoping evaluation for production models: what you optimize is what you get, and a badly shaped reward teaches the wrong behavior fast.