Knowledge Graph — Coursera Notes › Academic disciplines › Information Technology / Computer Science › Artificial Intelligence › Machine Learning & Data

Reinforcement Learning

concept · part of Machine Learning & Data

A machine learning paradigm where an agent learns a policy through trial and error to maximize cumulative reward from an environment.

Key algorithms include Q-learning, policy gradients, and temporal-difference learning, as detailed in Sutton and Barto's definitive textbook.

In the simulation, robots receive rewards for successful task completion and penalties for failures, improving their navigation and coordination over time.

A reward function provides feedback to an agent based on its actions. In the warehouse simulation, robots are rewarded for completing tasks without conflict and penalized for failures, enabling learning.

def reward_robot(robot, task_completed):
    if task_completed:
        robot.reward += 10
    else:
        robot.reward -= 5

Inside Reinforcement Learning (9)

• Average reward per episode — Cumulative reward divided by number of episodes, smoothing fluctuations to reveal learning trends.
• Cumulative reward — Total rewards collected in an episode or across episodes, often discounted by factor γ.
• Episode length — Number of steps taken to complete an episode; decreasing length may indicate more efficient goal-reaching.
• Exploration vs. exploitation ratio — Balance between trying new actions and using known high-reward ones, e.g., ε-greedy with ε=0.1.
• Policy stability — How often the agent changes its policy after convergence, indicating confidence and consistency.
• Reinforcement learning in production — Reinforcement learning is used by Netflix to optimize recommendation algorithms over time by maximizing user satisfaction and engagement.
• Sample efficiency — How well the agent learns from few experiences; critical when data collection is expensive.
• Success rate — Ratio of successful episodes to total episodes, a direct measure of task completion.
• Time to convergence — Number of episodes or steps until the policy stabilizes; lower is better when training is costly.

Connections

• Related to Supervised Learning
• Related to Unsupervised Learning

This is the text view of an interactive 3D knowledge graph — open this page with JavaScript enabled to explore it visually.