Multi-agent reinforcement learning versus multi-objective reinforcement learning

Question

everyone.

What is the difference between Multi-agent reinforcement learning and Multi-objective reinforcement learning?

And can you explain the pros and cons of the two methods?

I think both methods can serve some objectives...

Answer 1

In multi-agent reinforcement learning, many agents will take actions in the same environment, and each agent will try to maximize its own goal (which can also be the same as all the other agents).

On the other hand, multi-objective reinforcement learning has nothing to do with the number of agents, but rather on how the reward is designed: in multi-objective reinforcement learning the reward is simply a function of many goals that you are trying to accomplish (e.g. you want to achieve some generic task A, but with as less energy as possible)

Answer 2

Multiple-agents and multiple-objectives are orthogonal concepts. They can be combined together.

Examples of multiple-objectives:

• Optimizing for both space and time
• e.g. trading bot: Maximize profit while maintaining risk-neutrality
• self-driving car: Control steering, acceleration, and braking while keeping me safe and the ride comfortable. - Smoothness + safety
• human: Chew-gum and walk simultaneously
• robot: Walk with a human like gait, without wearing the joints too much, unless it's really important, then you can push it a little

Agents are independent decision-making systems, like individual players in a game. Multiple agents can cooperate or compete with each other, in either shared or separate environments.

Cooperatively, they seek to maximize at least two objectives, their individual performance while also maximizing the objective of the team's performance. This is a multi-objective problem in a multi-agent problem.

Most production systems have multiple objectives either implicitly or explicitly.