Main menu:

Site search




Monica's Other Sites

Intelligence Is For Prediction

This blog entry elaborates on previously published material.

The purpose of Intelligence is Prediction. Evolution of the ability to predict the behavior of other agents and the likelihood of phenomena in the environment improved survival rates and created a strong evolutionary pressure to develop better and longer term predictions. This is how and why Intelligence evolved.

Have you ever wondered why robots walk with a robotic gait? It is similar to a “zombie” gait, and as we all know, zombies have no brains. A running zombie would look like it was intelligent, wouldn’t it?

A running two legged robot would be very impressive. The engineers at Honda know this; their robot Asimo attempts to “run” but I think it still has a ways to go. Decide for yourself. To me, it looks as if it is running on eggshells, or that it cannot cope with the slightest variation from the smooth, level surface beneath its feet.

If you are happy to just walk and if you have four legs, then the walking gait can be spectacularly graceful, as in the case of Boston Dynamics’ Big Dog.

Imagine standing in somewhat complex terrain, such as on a rocky beach. Blindfold a friend, spin them around, and ask them to walk ten meters forward. Their gait will resemble that of robots. They move a foot forward, make sure the foot is in a stable position and can carry their weight before lifting the other foot. Their movements are based only on feedback from nerves sensing pressure on the soles of their feet, and balance sensors in their inner ear.

How might running gaits have evolved in animals, starting with animals living on the bottom of the ocean? As far as I know there are no bottom-dwelling walkers with fewer than six legs.

Four or more legs allow you to “take risks” when moving one or more, while trusting the remaining legs to keep you stable in very irregular terrain.

Multi-legged Walking Requires Central Control

The earliest brains might have been nothing more than simple clusters of nerves that coordinated the walking of legs in multi-legged agents, such as crabs, insects, or other arthropods.

Walkers keep some of their legs on the ground at all times, so you can only lift a few of them at any one time. You will also want to avoid stepping on your own feet. Subsumption architectures, which stratify behaviors according to the immediacy of their goals and importance of their fulfillment, allow multi-legged robots such as Rodney Brooks’ Attila to walk with something resembling the “confidence” of an animal, and, importantly, with an animal’s speed, both in locomotion and in error recovery. I could believe that centipedes and other animals with undulating gaits might well use distributed control but it seems that in nature, anything that uses legs in a non-undulating fashion controls the movements from a central “brain”.

The buoyancy provided by water means that crabs and other bottom-dwellers can worry less about whether their legs can carry their weight or not, and how secure a footing they need to maneuver. They can “run” underwater, in a pell-mell fashion. This also works on dry land as long as you have legs to spare, so momentarily losing solid support under some of them won’t stop you.

Certain land-based arthropods can move quite rapidly, and some can jump long distances. By the definition below, these jumpers might qualify as “runners”. But true running evolved in land based mammals, and those have only four or two legs.

The fewer legs you have, the more energy-efficient your gait is. Kangaroos use less energy per distance traveled when bounding along on their hind legs rather than when using 5-limbed locomotion (four legs and the tail).

Running Requires Prediction

When using a “running gait”, so many legs will be off the ground at the same time for some part of the running cycle that the remainder cannot carry the body weight. This is important, because it means you can no longer rely on feedback from a leg to tell you that the leg is well positioned on the ground, that the ground under the leg will carry its share of the body weight, and that it won’t slip. You must predict (with something like millisecond precision) the spot where the leg will land, and the impact moment, so that you can start preparing the appropriate muscles for the landing impact, rebalancing, and for the next step. You also need to predict future adjustments to body posture appropriate for dynamic balancing, for example to account for changes to the terrain as perceived by the eye.

All the while, of course, avoiding predators.

This kind of prediction would also improve your speed in a walking gait, which means there are continuous rewards for small improvements. Clearly, development of this predictive ability through Evolution is Biologically Plausible.

Let me quickly define what I mean by “Biological Plausibility”: In order for a feature to be Biologically Plausible there must exist a way that the feature could have evolved. This requires a starting configuration (such as the existence of walking animals) and a gradient of advantage that provides an evolutionary pressure to develop the feature. A continuous gradient where small changes yield small improvements is much preferred over steeper saltations where a major change provides a major improvement without intermediate steps.

Running Requires Vision

Now imagine running blindfolded on a rocky beach.

Vision evolved before running. Running requires vision so that you can run without hitting obstacles, predict where your feet can land, and plan ahead for a path that provides solid spots to place your feet.

What’s around the next bend

If you can remember features of places in the environment from one visit to the next you can remember safe and dangerous places to go, safe and treacherous spots to plant your feet, etc. You are capable of predicting what the environment looks like ahead when you move, in order to move faster.

Prediction is a Biologically Plausible Feature

There is a strong evolutionary pressure to get better and better at predicting your environment. You would benefit from developing behavioral models of other agents to predict how predators or prey will act, and you would want to predict the behaviors of other members of your tribe. such as potential mates and rivals. The single skill of prediction, even if it often fails, yields a big advantage in how well you survive and how likely you are to breed. The better you can predict the near future compared to your immediate rivals, the more offspring you will have.

Jeff Hawkins also believes Intelligence is defined by Prediction. You may enjoy this excellent video. He suggests that only higher levels of Intelligence use prediction. but I believe prediction is a fundamental low-level operation for all intelligent agents.

To summarize, the purpose of Intelligence is Prediction. The ability to make predictions is a Biologically Plausible feature of brains.

Related Reading

Distributed mechanical feedback in arthropods and robots simplifies control of rapid running on challenging terrain


Pingback from Essential Skills for 21st Century Survival: Part I: Pattern Recognition « emergent by design
Time April 5, 2010 at 9:31 am

[...] The ability to spot existing or emerging patterns is one of the most (if not the most) critical skills in intelligent decision making, though we’re mostly unaware that we do it all the time. Combining past experience, intuition, and common sense, the ability to recognize patterns gives us the ability to predict what will happen next with some degree of accuracy. The better able we are to predict what will happen, the more intelligent we become. So, you might say that the purpose of intelligence is prediction. [...]

Write a comment

You need to login to post comments!