The Adaptive Interaction principle was pioneered by American researchers Robert Brandt and Feng Lin in their publication titled “Adaptive interaction and its application to neural networks” in Information Sciences Journal in 1999.  In this work, the authors laid the foundation of a very exciting type of Artificial Intelligence systems based on dramatically different principles.

Typically, all Artificial Intelligence systems are built around Artificial Neural Networks (ANN) of various kinds that are trained by other algorithms, such as backpropagation, neuroevolution, and similar.  Adaptation in such systems is usually performed in the form of either supervised or unsupervised learning.  In order for such training to work, a designer needs to build a process model from control laws, then identify the unknown parameters that can affect the system and finally use some type of synthesis to adjust parameters of the model so the controller can adapt to changes.

Obviously, this is not how things adapt in the nature.  A fruit fly with its 2,000 neurons can’t possibly build a flight dynamics model of flapping wings system, account for unknown factors affecting it in flight and at the same time run visual and smell recognition algorithms and perform other tasks it usually performs.  We suspect that fighter jet pilots also do not build complex multi-dimensional mathematical models of their planes and then solve them in real time – they just don’t look the type. Therefore, everything points to the fact that the conventional approach to adaptation in artificial systems needs to change if they are to approach the sophistication of biological systems.

We need to get rid of rigid models, complex data fusion algorithms, and other such things, and use the methods that Mother Nature developed during billions of years of evolution. If we look at how natural systems solve complex problems, we will see the obvious – natural systems break down big problems into a myriad of very small and simple to solve problems. For example, to build a system to extract water from soil, then channel it through a pipe system tens of feet up, and distribute it to thousands of points of use would have taken human engineers years to design. Trees accomplish this task naturally and pretty efficiently without knowing anything about mechanical engineering, colloid physics and other similarly complex matters. This is because trees are composed of little cells that interact with each other. Each one performs a very simple task that doesn’t require much brain power to complete. By interacting with their neighbors, these cells can regulate water supply, change chemistry of a solution pumped through a tree trunk, and even shut down for the winter when temperature drops.

So, intuitively, this seems to be the way to build Artificial Intelligence systems as well. Unfortunately, no existing systems employ such an approach, and even those that attempt to are still subject to cruel and unusual punishment from their creators – the engineers. Engineers tend to be control freaks – they want to understand what the system will do at any given time, so they “improve” their AI systems with checks, limits, failsafe switches and other gizmos that look good in theory, but don’t work well in practice.

On the other hand, the authors of the original Adaptive Interaction theory also greatly oversimplified their system. They did not account for a few very important points that made their original algorithm all but useless in the real world.

This is where R3’s team took the Adaptive Interaction theory to a new level. We called it Dynamic Adaptive Interaction (DAI). Not only can DAI adapt to extremely complex non-linear processes with discontinuities, it will also alter its learning rate based on the environment and mimic natural systems. All adaptation is automatic and happens without any human intervention or prior knowledge of the system being controlled. R3 has developed a number of systems based on Dynamic Adaptive Interaction concept, including an autonomous self-learning flight controller for unmanned aerial systems, a self-adapting navigation sensor, autonomous VTOL landing system, a self-training servo actuator, and other equally unique and innovative solutions.