Multi-Agent Reinforcement Learning Basics Quiz Quiz

Explore fundamental concepts, cooperative dynamics, and essential terminology in multi-agent reinforcement learning with this beginner-friendly quiz. Designed for learners and enthusiasts, this quiz helps deepen understanding of how multiple intelligent agents interact and learn in shared environments.

1. Definition of Multi-Agent Reinforcement Learning

   What best describes multi-agent reinforcement learning in artificial intelligence?

   1. A method where agents use supervised learning only
   2. A system in which agents operate completely without any type of feedback
   3. A process where multiple intelligent agents learn and act in a shared environment
   4. A type of task where only one agent interacts with the environment independently

   Explanation: Multi-agent reinforcement learning involves several agents learning and making decisions within the same environment, often interacting with each other. Single-agent scenarios involve only one agent, not multiple. Supervised learning is a different paradigm where correct outputs are given, as opposed to agents learning from rewards. Agents without feedback are not practicing reinforcement learning.

2. Characteristics of the Environment

   In a multi-agent setting, what often changes about the environment from an agent's perspective compared to a single-agent setting?

   1. The environment becomes non-stationary due to the actions of other agents
   2. The environment stays completely stationary at all times
   3. The number of actions available to an agent decreases
   4. The reward function is always negative

   Explanation: In multi-agent systems, the environment can appear to change unpredictably because other agents are also learning and changing their behavior. Unlike in stationary environments, the transition dynamics may shift as agents adapt. There is no rule saying the reward function must be negative. The number of possible actions usually remains the same or increases rather than always decreasing.

3. Types of Agent Relationships

   Which term describes agents working together to achieve a shared goal, such as moving a box together?

   1. Antagonistic
   2. Isolated
   3. Cooperative
   4. Competitive

   Explanation: Agents with a shared goal, like moving a box, display cooperative behavior where success depends on collaboration. Competitive agents are trying to outdo each other, not cooperate. Isolated agents do not interact, while antagonistic is not a commonly used technical term in this context.

4. Credit Assignment Problem

   What is the main challenge known as the credit assignment problem in cooperative multi-agent reinforcement learning?

   1. Deciding the next possible state
   2. Assigning negative rewards only
   3. Setting an agent's learning rate automatically
   4. Determining which agent's actions contributed to the team's success or failure

   Explanation: Credit assignment refers to figuring out how much each agent helped or hurt overall outcomes, especially when rewards are shared. Assigning only negative rewards or setting learning rates are unrelated to credit assignment. Deciding next states is a result of policies and environment transitions, not directly credit assignment.

5. Centralized vs. Decentralized Training

   Which approach allows agents in multi-agent reinforcement learning to use global information during training, but only local information while acting?

   1. Reward sharing
   2. Single-agent planning
   3. Centralized training with decentralized execution
   4. Purely centralized execution

   Explanation: Centralized training with decentralized execution enables agents to share information during training, improving learning, but requires them to make decisions based only on their own observations during action. Purely centralized execution is rare in real-world applications. Reward sharing describes how rewards are distributed, not the training approach, and single-agent planning ignores other agents.

6. Partial Observability

   In many multi-agent environments such as hide-and-seek, why do agents often have only limited information about the entire environment?

   1. Because the setting is partially observable for each agent
   2. Due to guaranteed full observability for all agents
   3. As a result of fixed, unchanging rewards
   4. Because agents cannot take any observations

   Explanation: Agents may have only a local view and thus operate under partial observability, leading to decisions based on incomplete information. Full observability rarely applies in complex, multi-agent environments. Saying agents cannot take observations at all is incorrect. Fixed rewards do not relate to observability.

7. Competitive Multi-Agent Scenarios

   Which example best illustrates a competitive multi-agent environment?

   1. Two agents playing against each other in a chess game
   2. Several agents each moving randomly without a goal
   3. Multiple agents building a house together
   4. A single agent learning to walk

   Explanation: A chess game features two agents (players) each aiming to win, hence displaying competition. A single agent learning a task is not multi-agent. Random movement without a goal may not involve interaction. Building a house together is a cooperative scenario.

8. Communication in Multi-Agent Systems

   Why might communication between agents improve performance in a cooperative multi-agent task?

   1. Because it helps agents coordinate their actions and share observations
   2. Because it increases randomness in the environment
   3. Because it prevents all reward functions
   4. Because it limits the agents' possible actions

   Explanation: Communication lets agents share information, improving coordination and team performance in cooperation-based tasks. Increasing randomness or limiting actions do not enhance cooperation. Preventing all reward functions makes learning impossible.

9. Exploration in Multi-Agent Settings

   What is a common challenge with exploration in multi-agent reinforcement learning?

   1. Exploration is fully guided by an expert
   2. The presence of multiple agents can cause more unpredictable outcomes during exploration
   3. Each agent always knows the optimal policy
   4. There is never any variability in outcomes

   Explanation: When many agents are exploring at once, their actions affect one another, making outcomes less predictable and exploration more complex. It is incorrect to say outcomes never vary. Agents often explore independently, not guided by an expert. Agents rarely know the optimal policy during learning.

10. Decentralized Policies

    What does it mean for an agent to follow a decentralized policy in multi-agent reinforcement learning?

    1. The agent waits for centralized commands before acting
    2. The agent makes decisions based only on its own observations without access to global state
    3. The agent makes decisions using all information from every other agent's perspective
    4. The agent always uses a single, fixed action

    Explanation: A decentralized policy means the agent acts based on local information, which is realistic in many applications. Relying on the global state is the opposite of decentralization. Using a fixed action ignores learning and adaptability, and waiting for central commands is not acting independently.