Multi agent Control Architectures

What is an agent?

Most people define an agent as something that continuously executes a loop:

loop forever:

• Sense
• Decide
• Act

The agent senses its environment, and based on what it sees and also its model of the world, it makes a decision as to what to do next, and then does it.

Approaches for multi-agent coordination

If there are many agents, as in our simulation, there is usually a need for a mechanism for agent coordination.

There are three approaches to decision-making for multi-agent coordination: centralized, distributed and hybrid. In a centralized approach, all coordination is done by a single system. In a distributed approach, all coordination is performed by the agents themselves, with no central decision-maker. In a hybrid approach, coordination is partly done by a central system and partly by the agents themselves.

Pilot-controller decision-making for coordination

Together, the pilot and operators in the control tower define a hybrid system for multi-agent coordination. We have been discussing recently how the decision making is divided between the pilots and the controllers, especially when it comes to what to do when something unexpected happens. At first, we thought the pilot agents should be responsible for detecting something unexpected (like another aircraft approaching an intersection when it is doing so as well) and should decide to slow down or stop. But today we discussed that this might not be correct. The detection of something unusual should be done by a central control system (a controller agent, not a pilot agent). We therefore assume that the controller agent has a larger, global view of the airport surface, and that this agent can tell when something on the surface is different from what it should be (an aircraft should be at location A but instead it is at location B). From the controller's global 'sensor' it can decide what the other aircrafts on the surface should do in response to this anomaly. This decision is communicated to the pilot agents who then collectively act (slow down or stop) to avoid any potential unsafe conditions.

Notice that on this model we have taken some of the decision making of the pilot and moved it to the controller. As we discussed, this is probably a more realistic model of what actually happens.