Policy Gradient CartPole — Evolution Strategies

A classic reinforcement-learning benchmark, trained live: a linear policy network reads the pole's angle and the cart's position/velocity and outputs a push force, and instead of hand-coded rules or table-based Q-learning it is trained purely from trial and reward using Evolution Strategies — sampling weight perturbations, running an episode with each, and nudging the weights toward whichever perturbations balanced the pole longest. Watch the pole go from falling almost immediately to staying upright for the full episode over a few dozen generations, adjust the learning rate and exploration noise to see the classic RL stability/exploration trade-off, and read the actual weight vector and gradient-update formula driving it.