19 March 2009
Autonomous-agent adaptive systems
By Jim Pinto
Smart, wireless sensor networks will soon be everywhere, collecting and transmitting vast amounts of previously unmeasured data. To accommodate higher I/O point count, significant new advantages and benefits will emerge through autonomous-agent adaptive systems.
At the input/output level, most of today's systems are similar: Clumps of I/O connected via a device network to a deterministic, hierarchical system, which is prone to failure when complexity increases.
Conversely, interacting autonomous agent systems provide predictable performance-locally non-deterministic, but globally stable and predictable.
Conventional control systems are deterministic because non-determinism is considered unacceptable. A hierarchical control system is subject to complex software constraints, communication conflicts, and network-speed limitations. The upper limits appear to be constrained only by data rates. The system solution of increasing bandwidth does not work. Removing the need for data rate does.
Few contemporary control systems have achieved practical I/O point counts of more than a few hundred thousand. By contrast, programmable, intelligent, autonomous I/O systems with algorithmic (rule-based) response mechanisms have no theoretical complexity limit. Vastly improved performance will be achieved at a fraction of the cost of deterministic hierarchical systems.
Over the past several years, Complexity (Chaos) Theory analysis is demonstrating deterministic systems are, in reality, non-deterministic (brittle or prone to failure) when complexity increases. Conversely, it is being demonstrated that interacting autonomous agent systems provide predictable performance-locally non-deterministic, but globally stable and predictable.
By "autonomous" I mean an intelligent, rule-based I/O cluster, with minimal instructions in each node, managed remotely by the network. Autonomous Agent architecture is where the next major software advances will occur. This will bring significant price/performance improvements by putting the processing right at the front-end-at the sensors and actuators. When individual devices become programmable and re-programmable over the network, the ability to self-organize becomes inherent. The "agents" are not programmed; the "behavior" is modified.
The software architectures required to achieve "self-organizing systems" are still somewhat new, but practical applications are emerging. Change in the relative relationships among cost/size/complexity heralds the next revolution in control. The transition will occur away from traditional hierarchical and procedural controls to rule-based behavior.
Beyond just robustness of the system and improved overall performance, the increasing connectedness of peer-to-peer control systems brings Metcalfe's Law (exponentially increasing effectiveness) and the new science of complexity into play. The result is an intrinsically different kind of operation: Complex adaptive systems. These systems yield significant advances through reduced software complexity, faster and easier installation, robust performance, and vastly improved flexibility, with the capability to handle much larger numbers of real-time sensing and control connections.
These types of systems have roots in the work being done on artificial life and genetic algorithms at places like the Santa Fe Institute. During the next decade, the results will become evident in factory automation and process controls. If you wait, you will be too late.
Behind the byline
Jim Pinto is an industry analyst and founder of Action Instruments. You can e-mail him at firstname.lastname@example.org or view his writings at www.JimPinto.com. Read the Table of Contents of his book, Pinto's Points, at www.jimpinto.com/writings/points.html.