I found a nice example of the use of an agent-oriented approach over at the
AOS website. The example describes an agent-based aircraft control system. Firstly, the problems are depicted, identifying:
- Chaos from many aircraft converging on an airport
- The need to hold aircraft on approach to airports
Even in a medium-sized airport, there will be congestion at certain times of the day, with more aircraft arriving than can be accommodated by the runway(s). Resolving this competition for the runway resource requires that the flow manager take many factors into account, and can lead to overload:
- Too many unexpected dynamic changes in the situation
- sudden change in wind
- aircraft diverging from expected behaviour
- one of the runways unavailable due to weather
- Too many aircraft in a given time frame
- too many ETAs (Estimated Time of Arrival) and pilot instructions to calculate/issue
- Optimisation factors
- potential for simultaneous landing on crossing runways
- heavy aircraft can land close behind light aircraft (but not vice versa)
- Leads to
- communications bottlenecks with air traffic controllers
- aircraft unnecessarily forced to enter holding pattern
- sub-optimal use of the runway resource
It is suggested that too much complexity and interaction leads to a need for delegation, one of the trends that Wooldridge identified as leading towards agent-oriented approaches. Furthermore
AOS identify that traditional software techniques are inappropriate because they are unwieldy. Although there is a single algorithm that could satisfy the problem, it is inefficient and doesn't deal nicely with the dynamism of the real world, hence it is a fragile approach:
The key point is that there is a single locus of control. The main loop grows with the complexity of the problem, and reactivity has to be manually incorporated into each step of the loop. Because of the inherent complexity of interactions, traditional software approaches lead to a monolithic code-base that is error-prone and difficult to alter without breaking the carefully coded handling of the dynamic aspects of the problem.
The agent-oriented approach is to delegate various functions to autonomous agents, essentially distributing functionality throughout the system and avoiding the centralised control loop. More on the proposed AOS solution, showing the types of agents required and relationships between them, is
here.
