Saturday, November 28, 2015

Robotic Fitness

Fitness has one meaning in the context of evolution, specifically, those qualities that assist an organism in replicating itself. For a bacteria, they include the ability to find food, perhaps moving to locations where it is, the ability to utilize something as food, either many things for an omnivore or unique things no other type of creature wants for a stenovore. Apologies but there is no word for the opposite of omnivore with the meaning I needed, so I made one up from an appropriate prefix and the same suffix.

A bacteria also has to be able to fend off predators or evade them or at least hide from them, or perhaps just reproduce so much that losses can be tolerated. It has to be able to survive infection, or take some actions to reduce losses from it. It has to be able to bud and form a twin. All but the last of these relate to the survival of the bacteria and the last is simply restating replication.

Similar fitness attributes, some almost identical, can be applied to every organism. These are the attributes that allow some species or subspecies to exist for a long period of time. They are the attributes favored by evolution, and also play a role in the competition between mutations to see which one will exist for the next long period of time.

There are two ways in which these can be emphasized. One is for situations where it is hard to survive, and then the survival skills are the ones selected by evolution. The other is for situations where it is easy to survive, and then the reproductive skills are the ones selected by evolution. These two classes of situations will lead to a selection, via evolution, of different and possibly contrasting choices for attributes. Survivability and fertility don’t select the same attributes, and organisms in these two different classes of situations will evolve differently.

Now let’s jump to a different environment, one in which an alien civilization has invented robots. The robots have a high degree of robotic intelligence owing to the aliens’ success in electronics, optical computing, algorithms, and other things they use to make robot brains. You could say that successive generations of robots are evolving, and there is a robotic fitness test that is making the selection of modifications. Mutations do not come from oxidation reactions in cellular milieu, nor from cosmic rays tampering with DNA, but from alien engineers tinkering with their products. Neither is the fitness test one of survival in the wild, nor of the ability to build another robot identical with the first one. It is a fitness test that measures how well the robot serves the purposes for which it is being used. Does it do the job it was created for, and what are the discrepancies, the costs, and the side effects?

If it is a single purpose robot, the metrics for measuring performance are fairly straightforward, but if it is a multi-purpose robot, some averaging over potential uses needs to be done. It takes some time to measure these things, but then the next generation of robots can be designed. The metrics certainly have an impact, but so also do coincident improvements in components that can be used to construct the robot. Software improvements go on at the same time. If algorithms are insufficient to make the robot mutations more successful, some machine learning can be employed. These involve the adjustment of parameters used in networks that make choices, or in networks that control motion, or other networks. The parameters influence how the smallest tasks are accomplished, for example grasping, and then layers are built up to coordinate small tasks into larger ones. The metric can appear at any of these levels, just as evolution would make selection based on both the accomplishment of timy tasks, like grasping, and also larger ones, such as releasing the grasp in a spear throw.

Machine learning is done against some metric. For an evolving creature, it is either survival or reproduction. For a robot, fitness means how well it accomplishes the tasks for which it was designed according to some set of metrics that the designers create. Conceivably, design could be automated and done by a robotic factory component, which would be informed of the tasks and the metrics. Perhaps it could obtain sub-metrics by observing successive generations of robots performing tasks, and using some reasoning to pick sub-metrics which contribute to the overall metric for this class of robot.

In this context, how does a robotic overlord arise who takes over the alien planet and subjugates its creators? This is a theme which appears in science fiction, and elsewhere. Does it happen by accident, where the alien engineers design the robot to do one thing, like wash dishes, and instead it does something else, like learn kung fu with which it kills its designers? Does somebody in the production process make a mistake, like putting a minus where a plus is supposed to be, and a robotic car takes off to hit as many people as it can instead of missing them all by a safe distance? Sign errors happen, of course, but so does testing, verification and validation, and checking. Does the error occur at a high level, say in the master controller of the water supply, and instead of flowing the right amount of water out of the reservoirs into the dwellings of the aliens, it puts poison into it, which it just happens to have handy because it also runs the factory making poison?

Perhaps it happens because something out of the ordinary happens, like a hurricane, which somehow taxes the robotic controller of the water supply so much that it starts drying up the city and then refuses to allow water to flow to the aliens, leaving them all to die of thirst. Or it might be more subtle, like a traffic controller that causes train crashes, killing all key programmers who understand the traffic controller system well enough to stop it.

There is no end to the scenarios that can be thought up wherein a robot of some ilk goes haywire and causes damage or death. Given an alien world where they are prevalent, that might happen. But the response to such an incident is to change the software that allowed it to happen. Imagining it to be self-generated, where the robot replaces the corrected software with the original faulty software is also possible, but all this seems so outlandish and unlikely, that it might be relegated to fantasy.

Robotic fitness is not a fantasy idea, it is the basis upon which robotics would be designed. Such a standard would not admit of any homicidal tendencies, nor any accidental problems for more than a single or a few instances, long enough to find the problem and repair it. Thus the concept of a world of mechanical beings can be understood, but only as a voluntary choice by an alien population. And then the question is why. It remains unanswered.

No comments:

Post a Comment