Machines that think – Robots that see
How might it be possible to build an intelligent machine? A machine such as, for
example, a domestic robot that could tidy up and know the difference between toys and
rubbish?
In our descriptions of machines we often use words that are used in everyday language for characteristics of animals and
humans, and mean them in a sense that applies to the characteristics of a program or a
device. For example, we say that a clock “goes”, not that it functions. This does not imply that we think animals and humans are
machines, but we do it to avoid a highly technical style of speaking.
When we build a machine we use animal and human behaviour as a source of ideas.
The philosopher Daniel Dennett distinguishes four levels of animal behaviour. He has named these levels after four
scientists: Darwin (1809-1882), Skinner (1904-1990), Popper (1902-1994) and Gregory (1923-).
- Darwinian creatures are characterised by fixed, instinctive behaviour. They often appear to be intelligent, but they cannot adjust when circumstances change. The biological example is an insect.
- Skinnerian creatures try different actions and learn from the results. A pigeon is an example.
- A Popperian creature assesses its options before performing an action. A biological example is a rat.
- A Gregorian creature uses mental tools. Humans are Gregorian creatures.
If we use Dennett’s categories to classify machines, a machine in the ordinary sense would be a Darwin machine – a device that serves a specific
purpose. It would typically be gradually improved over time, as our experience of using it
increased. The goose quill, used to write with in former times, has now been improved to a modern ballpoint pen. We can think of this type of improvement as being analogous to the way in which evolution changes the bodies and behaviour of
animals.
A Skinner machine is able to respond when circumstances change. The prototype is a
thermostat, which adjusts the heat supply according to the temperature.
An example of a Popper machine is “Deep Blue”, the computer which succeeded in beating the World Champion at
chess. The method it used was, precisely, to foresee the result of an astronomical number of sequences of moves and then make the move that resulted in the most favourable position. As yet there are no Gregory
machines.
Research at the University of Copenhagen
Researchers at the University’s Department of Computer Science (DIKU) and researchers at the IT University of Copenhagen are in process of constructing a universal Skinner
machine.
“Universal”, in the sense that the user can set it to the application
desired. The machine will be equipped with remote control, and the user will guide it through examples of the desired
function. After that, the machine will perform the function on its own.
As it gains experience, it will adapt the way it works. It will learn from
experience.
We are designing the machine so that its behaviour will be based on its experience of life
(cf. our introductory remark about everyday and technical language). In this we are following Ernst Mach’s thesis that it should be possible to understand a person’s behaviour on the basis of the sensory signals that that person has received in the course of his or her
life.
For the body of the machine we are using a minirobot consisting of a small platform measuring about 10 cm each way and provided with
wheels. The platform will carry a computer, a camera and three rangefinders.
We will teach the robot very elementary behaviour. It must learn not to collide with
objects. It must learn to follow or avoid another robot. It must learn to push an object in a specific
direction. See the robot here (download RealPlayer here).
It would be possible to make the robot perform one of these functions by giving it explicit
instructions, in other words, by programming it to react correctly in every thinkable situation.
Instead, we will lead the robot through examples of correct behaviour, and then leave it to find out for itself how to act sensibly in new situations.
Challenges
There are several challenges in proceeding in this way. When explicit programming is replaced with generalisation from
examples, one may ask how reliably the robot will actually perform its task. How many examples must be given to achieve a certain
reliability? Is it at all acceptable to have a machine that is not completely
reliable? Will not the data processing be such a large task that it will be impossible to get the machine to work within a reasonable time?
Our Skinner machine will build up a simple model of its accumulated experience from various situations. We are planning to use a simple principle to extend the design to a Popper
machine. According to Dennett, a Popperian creature uses its body as a “sounding board”. The animal imagines an action and if its body reacts with aversion, it does not perform the action. Using this
analogy, we will construct a Popper machine that performs a “Deep Blue”-style simulation of the characteristics of actions.
Is it possible to compute such a model?
The answer is a qualified yes. Our basic thesis is that the I/O sequence contains the necessary information. A reservation has to be made because of the fact that we have to convert the sensory signals into data, into a finite number of bits. It may be that we lose crucial information because of this
requirement. Therefore it may well be that our model will not be precise. We have to estimate how frequently we need to capture data and how many bits to
use. If we guess wrongly, the model will not fit.
This requirement is critical, because the future behaviour of a chaotic system will be altered in an unpredictable way if the initial conditions are in the slightest
different. However, given that we have the necessary bits available, and that the world the robot moves in is not
chaotic, the task is “just” a problem that a computer scientist can solve.
Our aim is to build a robot that mentally “plays chess” on the basis of a universal entropic principle in order to decide on its next action. In other
words, the robot will follow a strategy based on assuming that the surrounding world is not more complex than there is evidence for in its
experience. At any point in time, it will postulate and maintain the simplest model of the world around
it.
The robot must be able to be set to solve a new problem using examples of the solution. It must be able to accumulate its experience and learn from
it.
Relevant literature:
E. Mach: Analysis of
Sensations, Dover Edition, 1962, Original published in German, 1886.
Daniel C Dennett: Kinds of Minds, 1996.^
|
|
|
|
|
|
 |
 |