The frame problem of artificial intelligence is discussed by Daniel C. Dennett in "Minds, Machines and Evolution: Philosophical Studies [Dennett 1 984]" by reference to a robot named R1 whose only task was to fend for itself. The frame
problem is one of deciding what items to consider when trying to predict the outcome of some action. Given finite resources there is not time to consider everything. The difficulty is in deciding which consequences need to be evaluated. This appears to be a problem which must be dealt with by any creature which is to be considered intelligent and is to prosper in a possibly hostile and slowly changing environment.
The example of the frame problem used by Dennett is as follows:
Once upon a time there was a robot, named R1 by its creators. Its only task was to fend for itself One day its designers arranged for it to learn that its spare battery, its precious energy supply, was locked in a room with a time bomb set to go off soon. R1 located the room, and the key t o the door, and formulated a plan to rescue its battery. There was a wagon in the room, and the battery was on the wagon, and R1 hypothesised that a certain action, which it called PULLOUT(WAGON,ROOM) would result in the battery being removed from the room. Straightway it acted, and did succeed in getting the battery out of the room before the bomb went off. Unfortunately, however, the bomb was also on the wagon. R1 knew the bomb was on
the wagon and could have deduced that moving the trolley would m ove the bomb. It had not been programmed to consider the side effects of its actions.
Back to the drawing board. 'The solution zs obvious, ' said the designers. 'Our next robot must be made to recognise not just the intended implications of its actions, but also the implications about their side effects. ' The new robot R2D2 (robot deducer) was placed in the same predicament. It too hit upon the idea of P ULLOUT(WAGON,ROOM), it began, as designed, to consider the implications of this act. It had just finished deducing that pulling the wagon out of the room would not change the colour of the room 's walls, and was embarking on a proof of the further implication that
Chapter 4 Neural Networks and Intelligent Systems 4.9
pulling the wagon would cause its wheels to turn more revolutions than there were wheels on the wagon. . . when the bomb exploded.
The frame problem is one of deciding what is relevant in a particular situation and has proved very difficult to overcome [Dreyfus 1 993] .
If we look at the assumptions underlying the description of the frame problem they seem to include:
a) A system that could use logical sequential thought should be considered intelligent. By "use logical sequential thought" we mean that it has knowledge of its environment and possible actions, a set of rules giving results of certain actions and a method of combining these to find the implications of proposed actions.
b) An intelligent being (the robot) should be able to look after itself, in this case avoid being destroyed by the bomb.
These two assumptions lead to the frame problem. What grounds do we have for either of these assumptions? If we accept them then the R 1 example clearly shows an intelligent being which cannot look after itself, but intelligent beings should be able to look after themselves, we are intelligent and we can look after ourselves, hence the paradox. Both of the assumptions are suspect but the key point is that we have no reason to assume that a faculty for logical sequential thought by itself is a survival trait. As an additional feature, used with the other abilities, it certainly has benefits but by itself it may be of no value. Even in conjunction with the level two abilities described previously, too much LST may be detrimental to survival. It could even be argued that too much dependence on LST must be detrimental to survival otherwise people would be better at it! The paradox of the frame problem, like many other paradoxes, may merely be the result of faulty assumptions. It is possible to avoid problems with the first assumption (LST = intelligence) by regarding it as a definition of intelligence. I suggest this is a poor definition of intelligence, it merely devalues the word to match a method which it is hoped might produce intelligent behaviour. Even if we accept this definition the paradox of the frame problem still disappears. The question just becomes why should intelligence be sufficient to provide survival. Both assumptions are required to form the paradox.
Intelligence is a difficult word to define with definitions rangmg from "That which is measured by intelligence tests " to more vague but in some ways more
Chapter 4 Neural Networks and Intelligent Systems 4. 1 0
satisfying definitions such as "General problem solving ability" or "Ability to survive in a complex world" [see Sternberg & Detterman 1 986 or Boring 1 923] .
It is perhaps the difficulty of defining intelligence that has led to the formulation of the frame problem. When we fmd a word difficult to define it is easy to fall into the trap of choosing a simple definition (in terms of its formulation) but then using the word as though it still had the more generally accepted though ill defined meaning.
The human ability to survive in a complex changing world, where decisions
must of necessity be made quickly and without full information may well
depend on the "older2" parts of the brain which do not use LST. Conditioned neural networks which have been trained by repeated exposure to the real world do not need to decide which items are relevant, all past events are used implicitly in deciding actions. All past events are included in the network to the extent that the connection strengths between neurons have been shaped by expenence.
It is true that we occasionally use a small amount of LST to decide on an action but we may well do this at the risk of falling into the frame problem. It has been noticed at least since Shakespeare's time
(if
not before) that those members of mankind who are too good at LST may not be good survival material.And thus the native hue of resolution
Is sicklied o 'er with the pale cast of thought, And enterprises of great pith and moment with this regard their currents turn awry And lose the name of action
( Hamlet,/11, i,56)
The existence and success of expert systems appears, at least initially, to be inconsistent with this theo
}y
of the purpose of LST. This issue is considered in the next section (expert systems).2 The term older is used in an evolutionary sense to refer to parts of the brain which evolved in the more distant past. A simplified description of this form is given by Ornstein & Thompson [Ornstein & Thompson 1986] .
Chapter 4 Neural Networks and Intelligent Systems 4. 1 1