RL Research Portfolio

A self-directed research portfolio (github.com/Amo-Gideon/rl-portfolio) exploring decision-making from tabular RL to world-model planning:

• Reproduced Dueling/Double DQN from scratch, with ablations on target-network updates and exploration schedules.
• Built an end-to-end RLHF pipeline (SFT, reward model, PPO alignment) for a 0.5B-parameter language model.
• On a PyBullet 2-link robot arm, PPO reaches 82% single-target success while a learned world model (MLP ensemble) with CEM planning reaches 100%, and stays robust under out-of-distribution physics shifts.
• Obstacle avoidance as a CEM cost term cut collision rates from 57% to 23% in a setting where five PPO training configurations failed.