What are MAS?

Multi Agent Systems (MAS) are an interesting subject in the context of AI. It can also be assigned to the more familiar swarm intelligence, which allows us a more descriptive view.

Imagine a colony of ants.
This colony (as a whole) is extremely efficient in its way of existence with all related, complex and superior tasks. However, if we look at the individual ants and their specific tasks these are less complex and hardly allow to draw conclusions about the overall tasks of the colony.

This is where MAS come in.

MAS essentially describe a coalition of several individual agents into an integral problem-solving system.

Individual agents are characterized by the fact that they interact autonomously with their environment and beyond that they can perceive their environment in some way.
The decision regarding which of the different interactions to perform is made by predefined goals that an agent should achieve; thus, it has a task.

In both, robotics and software development attempts are made to keep the individual agents as simple as possible from a technical point of view so that their behavior can be easily interpreted (similar to the individual ant).Therefore, the problem-solving efficiency of a single agent is also limited and is only generated by the interaction with other agents.

On the one hand, agents of the same type can be used for the MAS (all agents do the same thing in principle), on the other hand, different agent types can be involved (all three criteria could be affected here: Different goals, different interaction possibilities or different environment perception). Looking at communication (and interaction) with each other, this is often done via one-to-one or many-to-many communication, but you can also use extra agents that are used specifically for communication (one-to-many).

In principle, agents can be provided with simple logical rules, but it becomes even more interesting when they are complemented by machine learning methods that allow for higher dynamics.
Here, several possible learning methods arise, which all aim at improving the overall system, but either relate to the training of each individual agent itself or perform higher-level training that affects a group of agents.


When to use MAS?

Analogous to the logic of the Deep Neural Networks, it can be said that almost all problems can be mapped with Multi Agent Systems. However, it is not necessarily advantageous or advisable to have simple problems solved by quite complex MAS structures. Vice versa, MAS are suitable for the (approximate) solution of e.g. NP-hard problems whose exact solution is difficult to compute or even impossible to compute in a given time period.

Currently, we are developing and testing MAS in spatio-temporal ordered supply chains.

The key to success

There are several things to consider. First of all we need a definite domain in which all and every agent can operate. We have to choose an overall MAS architecture regarding types of agents, number of agents or communication types. This is followed by the interesting task of translation into mathematical formulas for objective functions and constraints. In practice, it is necessary to work on the tasks in several rounds until you reach the goal.


To be continued.