MARL Quizz

What is Mutli-agent Reinforcement Learning (MARL)?

• Reinforcement learning with no environment involvement.
• Multi-agent Reinforcement Learning (MARL) involves multiple agents learning and making decisions in a shared environment.
• Agents competing without any interaction.
• Single-agent learning in isolation.

What distinguishes "state" from "observation" in reinforcement learning?

• State is environment's true variables; observation is agent's perceptions.

• States are internal; observations are imperfect perceptions.

• • States and observations are interchangeable in RL.

• States are the agent's knowledge; observations are external.

What is the significance of the "reward" in MARL?

• Reward is used to punish the agent for incorrect actions.

• Reward is the feedback signal used to shape the learning process.

• Reward is the measure of the agent's knowledge about the environment.

• Reward is the indicator of the agent's confidence in its decisions.

An episode is a sequence of decisions that end when the agent decides he learned enough

• True

• False

Which of the following is a key challenge in MARL?

• Limited communication between agents
• Non-stationarity of the environment
• Static environment dynamics
• Single-agent decision making

What does the training scheme "Centralized Training Decentralized Execution (CTDE)" aim to achieve?

• 
  To centralize both training and execution processes.

• To limit agent performance by restricting training methods.

• To enhance communication between agents during execution.

• To improve agent performance through centralized training while enabling decentralized execution.

Which of the following algorithms is commonly used in MARL?

• A2C

• PPO
• DQN
• MAPPO

In the context of multi-agent reinforcement learning (MARL), how does difference reward shaping address the credit assignment problem, and what are the potential trade-offs associated with its use?

• 
  It calculates each agent's reward as the difference between the global reward and the reward without the agent's contribution, addressing credit assignment but increasing computational complexity.

• 
  It shapes rewards to emphasize exploration, helping agents avoid local optima but sacrificing convergence stability.

• 
  It uses a centralized critic to assign rewards based on global objectives, ensuring fairness but reducing scalability.

• 
  It assigns rewards proportional to each agent's actions while ignoring other agents' contributions, ensuring independent learning but increasing non-stationarity.

What is action masking in reinforcement learning?

• A method to guide the agent by dynamically disabling invalid or irrelevant actions during decision-making.

• Action masking is a technique used to enhance the exploration of an agent's action space.
• Action masking refers to the process of rewarding all actions differently in reinforcement learning.

• Action masking is a method to increase the number of actions available to an agent.

What is the Markov Game (MG) in MARL?

• A Markov Game (MG) is a multi-agent framework where agents interact in a shared environment, making decisions based on states and joint actions.
• A Markov Game (MG) is a type of board game played by multiple players.
• A Markov Game (MG) only applies to deterministic environments.
• A Markov Game (MG) is a single-agent framework with no interactions.