🧪 Reinforcement Learning Advanced Simulation

This simulation trains a reinforcement-learning agent on the classic Mountain Car control problem live, in your browser. A car in a valley must learn that it cannot drive straight up the goal hill — it needs to reverse up the opposite slope first to gain enough momentum. The agent learns entirely from trial-and-error reward signals using SARSA(λ) over a tiled approximation of the continuous position-velocity state space, with eligibility traces carrying credit for a reward back across the states that led to it. Behind the car, a live 3D bar field renders the agent's learned cost-to-go landscape, visibly flattening as training converges; adjustable learning rate, exploration and trace-decay sliders let you see how each hyperparameter reshapes how fast — and how reliably — the policy is learned.