Tuesday, April 17, 2012

How Orientation might work

The following presents an argument how Boyd’s Orientation phase in his OODA loop model may work given its physical constraints. Support for this position can be found in “The Timing of the Cognitive Cycle” by Madl, Baars and Franklin, below. In that paper they set out the Learning Intelligent Distribution Agent, LIDA model, which measures the actual timing of the cognitive process and closely follows the strategic path laid out by Boyd.

The first step is to recognize that both the OODA loop and the LIDA model are comprised of a series of discrete but not necessarily distinct cycles. In this sense they are a process of ongoing, never-ending reiterations of themselves. “The repeated cycle of perception, understanding and action selection is called a cognitive cycle… such cognitive cycles are the fundamental building blocks of human cognition: ‘cognitive atoms’.” Single OODA loops, like LIDA cycles, are strung together into longer loops, i.e. they become a process, to make detailed planning and actions possible.

Orientation begins when observations enter “into our sensing or observing window.” These observations come in the form of discrete frames of information over different sensory channels. These frames are gleaned by an active search of the environment controlled by genetics, cultural traditions, previous experience, opportunity and chance. The limitations of these factors are considerable. Because information in the environment is in constant flux more of it is missed, ignored or used and discarded then is made available to cognition. The key idea here is that the “window” has a temporal, i.e. energy as well as a spatial, i.e. objective side.

The Second Law of Thermodynamics is the dominant principle on the temporal side. Objects and events move asymmetrically, energy dissipates and unless new information is gathered entropy increases. The Second Law creates a ‘global sense of time’ that constrains the environment, sensation, perception and cognition. There is also psychological or local time; the ability to project plans into the future, remember the past and to “build snowmobiles.” We can influence this type of time by manipulating the matter, energy, events and concepts that move past and through us on global time.

Visual perception is our dominant energy source and there is no clear-cut process of receiving and processing data as mentioned above. Once information has been selected through an active search process it enters two parallel but interconnected pathways through the cortex used for processing. The dorsal stream receives information from peripheral, ambient vision. Its concern is action, where important (identified and selected) objects are, how to respond to them and leaves little or no memory trace. The ventral stream carries foveal, focused objects and information. This is LIDA’s ‘understanding’ side of perception, deciding what is important, planning and the maker of memories. Of the two, the dorsal stream has a faster processing speed. That action can outpace understanding was an evolutionary advantage.

The ventral and dorsal streams form a sophisticated command and control system. The ventral stream is slower but richer, more robust and creates lasting impressions. It answers the questions what is important, what shall I act on and what will I use to act on it? It creates intent. The dorsal stream is faster, more efficient and creates mostly disposable impressions. It controls intent. These two streams interact to “answer what can be seen as the only question there is: ‘What do I do next?’”

When the OODA proceeds directly from observation to action by way of implicit guidance & control, analysis and synthesis is not used. The “building snowmobiles” step is bypassed. This is the fastest route possible for taking action into the world. It’s a straight genetic path constrained and framed by culture and experience. In the LIDA model this path is implied by the direct connection between sensory memory and sensory-motor memory along the dorsal stream. The path through “Understanding” is skipped and the current action flows directly from subconscious processing of the previous action.

This position finds support in what is known as the perceptual moment, psychological present, short-term, sensory or working memory. It’s the brief period where action can be directed from a precognitive state, the ‘autopilot’ stage of learned reflexive behavior. It’s an interval where behavior is acted out without reflective thought, planning or the use of long-term memory.

The average length of the psychological present is three seconds; the presumed top end is five. The LIDA model establishes benchmarks for the shortest times through both the ventral, i.e. conscious and dorsal, i.e. unconscious streams. Together they equal one cognitive cycle or an OODA loop. Here is a comparison of the intervals between the elements in OODA loops and cognitive cycles.

LIDA ‘perception’ is the interval when observations enter OODA’s “sensing or observing window.” This interval takes 80-100ms. The stimulus being received splits along both the dorsal and ventral streams. If action can be taken at this point the OODA’s implicit command and control and the LIDA’s dorsal stream connection between sensory memory to sensory-motor memory takes over. If doubt or suspicion delay this action beyond the 100ms threshold the signal in the ventral stream begins move through ‘understanding’ i.e. analysis and synthesis. This interval takes 200-280ms, which includes perception. At that point the LIDA model calls for a conscious broadcast while the OODA produces a hypothesis. At this point both models call for an action selection to be made, a bifurcation point. Either an action is taken or the signal can be fed back through the understanding/orientation for further refinement. The action selection phase takes 60-110ms so an entire cognitive cycle or OODA loop will run 260-390ms.

Two factors are apparent. The fastest way to cycle through both the OODA and LIDA is to use the direct dorsal route from sensory memory to sensory-motor memory. It effectively cuts the time between observation and action almost in half. The problem with this method is that it’s dumb to small environmental perturbations, what Boyd points out in Suspicion from Destruction and Creation. This path lacks vision, foresight or planning relying on sensory memory for future actions. It works in the context of short-term events like those possible within the psychological moment. While these actions can, up to a point be trained for, Ashby’s Law of Requisite Variety proves that eventually the ventral stream signal will be needed. Command needs to frame the objective for action.

Which brings up the second factor. How the cognitive cycles and individual OODA loops need to string together to become a richer process. It’s not a matter of extending a single loop or cycle over seven, eight, nine seconds or even longer. The Second Law renders such a concept untenable. Ideas and concepts have a basis in physiological energy so are subject to entropy, the pressure of passing time and a changing environment. It takes a lot of work to string a single thought out over a long period. It’s easier, faster and produces a more robust process when parts of older, by milliseconds, concepts are discarded and replaced by better fitting ones. Create, combine and act with the smallest cognitive cycles possible. When in doubt and as opportunity exists, think ventrally but not too long. When planning include opportunities for dorsal actions and shared heuristic problem solving as often as possible. This should produce short, vigorous OODA loops comprised of the fewest, thinnest OODA cycles possible.


Link to “The Timing of the Cognitive Cycle” by Madl, Baars and Franklin, http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0014803.