Every-Visit Monte Carlo implementieren

Die Every-Visit-Monte-Carlo-Methode unterscheidet sich von der First-Visit-Variante dadurch, dass Werte jedes Mal aktualisiert werden, wenn ein Zustands-Aktions-Paar auftritt, und nicht nur beim ersten Auftreten. Dieser Ansatz liefert eine umfassende Bewertung der Policy, weil er alle verfügbaren Informationen aus den Episoden nutzt. Gleichzeitig kann er jedoch mehr Varianz in die Wertschätzungen bringen, da alle Stichproben einfließen – unabhängig davon, wann sie in der Episode auftreten. Deine Aufgabe ist es, die Implementierung der Funktion every_visit_mc() zu vervollständigen, die die Aktionswertfunktion Q über num_episodes Episoden schätzt.

Die Wörterbücher returns_sum und returns_count mit Zustands-Aktions-Paaren als Schlüsseln wurden bereits initialisiert und zusammen mit der Funktion generate_episode() für dich geladen.

Diese Übung ist Teil des Kurses

Reinforcement Learning mit Gymnasium in Python

Anleitung zur Übung

Erzeuge eine Episode mit der Funktion generate_episode().
Aktualisiere die Returns und deren Zähler für jedes Zustands-Aktions-Paar innerhalb einer Episode.
Berechne die geschätzten Q-Werte.

Interaktive Übung

Vervollständige den Beispielcode, um diese Übung erfolgreich abzuschließen.

Q = np.zeros((num_states, num_actions))
for i in range(100):
  # Generate an episode
  episode = ____
  # Update the returns and their counts
  for j, (state, action, reward) in ____:
    returns_sum[(state,  action)] += sum(____)
    returns_count[(state,  action)] += ____

# Update the Q-values for visited state-action pairs 
nonzero_counts = ____
Q[nonzero_counts] = ____
    
render_policy(get_policy())

Code bearbeiten und ausführen

Diese Übung ist Teil des Kurses

Reinforcement Learning mit Gymnasium in Python

Hohe SchwierigkeitSchwierigkeitsgrad

4.8+

Kurs kostenlos starten

Dive into the exciting world of Reinforcement Learning (RL) by exploring its foundational concepts, roles, and applications. Navigate through the RL framework, uncovering the agent-environment interaction. You'll also learn how to use the Gymnasium library to create environments, visualize states, and perform actions, thus gaining a practical foundation in RL concepts and applications.

Exercise 1: Fundamentals of reinforcement learning Exercise 2: What is Reinforcement Learning?Exercise 3: RL vs. other ML sub-domains Exercise 4: Scenarios for applying RL Exercise 5: Navigating the RL framework Exercise 6: RL interaction loop Exercise 7: Episodic and continuous RL tasks Exercise 8: Calculating discounted returns for agent strategies Exercise 9: Interacting with Gymnasium environments Exercise 10: Setting up a Mountain Car environment Exercise 11: Visualizing the Mountain Car Environment Exercise 12: Interacting with the Frozen Lake environment

Delve deeper into the world of RL focusing on model-based learning. Unravel the complexities of Markov Decision Processes (MDPs), understanding their essential components. Enhance your skill set by learning about policies and value functions. Gain expertise in policy optimization with policy iteration and value Iteration techniques.

Exercise 1: Markov Decision Processes Exercise 2: Custom Frozen Lake MDP components Exercise 3: Exploring state and action spaces Exercise 4: Transition probabilities and rewards Exercise 5: Policies and state-value functions Exercise 6: Defining a deterministic policy Exercise 7: Computing state-values for a policy Exercise 8: Comparing policies Exercise 9: Action-value functions Exercise 10: Computing Q-values Exercise 11: Improving a policy Exercise 12: Policy iteration and value iteration Exercise 13: Applying policy iteration for optimal policy Exercise 14: Implementing value iteration

Embark on a journey through the dynamic realm of Model-Free Learning in RL. Get introduced to to the foundational Monte Carlo methods, and apply first-visit and every-visit Monte Carlo prediction algorithms. Transition into the world of Temporal Difference Learning, exploring the SARSA algorithm. Finally, dive into the depths of Q-Learning, and analyze its convergence in challenging environments.

Exercise 1: Monte-Carlo-Methoden Exercise 2: Episodenerzeugung für Monte-Carlo-Methoden Exercise 3: First-Visit Monte Carlo implementieren Exercise 4: Every-Visit Monte Carlo implementieren

Aktuelle Übung

Exercise 5: Temporal-Differenzen-Lernen Exercise 6: Die SARSA-Aktualisierungsregel implementieren Exercise 7: 8x8 Frozen Lake mit SARSA lösen Exercise 8: Q-Learning Exercise 9: Q-Learning-Aktualisierungsregel implementieren Exercise 10: 8x8 Frozen Lake mit Q-Learning lösen Exercise 11: Richtlinie auf einem rutschigen Frozen Lake evaluieren

Dive into advanced strategies in Model-Free RL, focusing on enhancing decision-making algorithms. Learn about Expected SARSA for more accurate policy updates and Double Q-learning to mitigate overestimation bias. Explore the Exploration-Exploitation Tradeoff, mastering epsilon-greedy and epsilon-decay strategies for optimal action selection. Tackle the Multi-Armed Bandit Problem, applying strategies to solve decision-making challenges under uncertainty.

Exercise 1: Expected SARSA Exercise 2: Expected SARSA update rule Exercise 3: Applying Expected SARSA Exercise 4: Double Q-learning Exercise 5: Implementing double Q-learning update rule Exercise 6: Applying double Q-learning Exercise 7: Balancing exploration and exploitation Exercise 8: Defining epsilon-greedy function Exercise 9: Solving CliffWalking with epsilon greedy strategy Exercise 10: Solving CliffWalking with decayed epsilon-greedy strategy Exercise 11: Multi-armed bandits Exercise 12: Creating a multi-armed bandit Exercise 13: Solving a multi-armed bandit Exercise 14: Assessing convergence in a multi-armed bandit Exercise 15: Congratulations!