About LearnRLFast Team
Joe Chang
Turn rewards into intelligent behavior
Follow every choice to its consequence
Reinforcement learning starts with a simple idea: an agent makes a choice, something happens, and the consequence changes what it should do next. LearnRLFast follows that process from states, actions, rewards, and policies into value learning, exploration, temporal credit, Q-learning, policy gradients, actor-critic methods, self-play, and deep reinforcement learning. Small environments and observable behavior keep the algorithms connected to what the agent is actually learning. The goal is to understand how experience can become better decisions—and how repeated decisions can grow into increasingly capable behavior.
Experience behind the lessons
LearnRLFast Team
Joe Chang
Joe Chang brings more than 20 years of software engineering experience across startups, research, international banking and FinTech, entrepreneurship, and production systems. A physics graduate of New York University, he has taught more than 1,000 students across machine learning, deep learning, computer vision, reinforcement learning, and large language models. He has also worked with universities and government programs on AI engineering education.
- Think in choices and consequences. Learn to frame reinforcement-learning problems around what the agent observes, what it can do, what it receives, and what its decisions change.
- Watch learning change behavior. Use small environments, trajectories, plots, and policy changes to see how experience reshapes what an agent does.
- Build from simple agents to modern algorithms. Move from tabular value learning to function approximation, policy gradients, actor-critic methods, PPO, and deep reinforcement learning.
- Understand RL in LLM post-training. Connect policies, preferences, reward models, and policy optimization to the reinforcement-learning ideas used to shape modern language-model behavior.
- Connect RL to robotics and embodied AI. See how policies, control, continuous actions, simulation, and interaction come together when agents must act in physical or simulated worlds.
- Explore self-play and multi-agent learning. Understand how behavior changes when agents cooperate, compete, communicate, adapt to opponents, and become part of one another's environment.
- Understand delayed consequences. Learn why an action can matter many steps later and how value functions, returns, and temporal credit connect decisions across time.
- Explore with purpose. Learn how agents balance what already works with actions that reveal better strategies, useful information, and unexplored possibilities.
- Learn from demonstrations and preferences. See how imitation, behavior cloning, preference data, and learned rewards can guide an agent when hand-written objectives are not enough.
- Shape behavior through rewards. Learn how reward design creates incentives and how better objectives can guide an agent toward the behavior you actually want.
- Measure real improvement. Compare policies across repeated episodes, seeds, baselines, learning curves, generalization tests, and task outcomes to establish whether learning is genuinely progressing.
- Diagnose learning dynamics. Recognize sparse rewards, unstable updates, overestimated values, weak exploration, model error, and brittle policies—and use those signals to choose the next experiment.