Introduction

 

Interesting work is being done in programming computers to solve problems which require a high degree of intelligence in humans. However, certain elementary verbal reasoning processes so simple that they can be carried out by any non-feeble minded human have yet to be simulated by machine programs.

This paper will discuss programs to manipulate in a suitable formal language (most likely a part of the predicate calculus) common instrumental statements. The basic program will draw immediate conclusions from a list of premises. These conclusions will be either declarative or imperative sentences. When an imperative sentence is deduced the program takes a corresponding action. These actions may include printing sentences, moving sentences on lists, and reinitiating the basic deduction process on these lists.

Facilities will be provided for communication with humans in the system via manual intervention and display devices connected to the computer.

The advice taker is a proposed program for solving problems by manipulating sentences in formal languages. The main difference between it and other programs or proposed programs for manipulating formal languages (the Logic Theory Machine of Newell, Simon and Shaw and the Geometry Program of Gelernter) is that in the previous programs the formal system was the subject matter but the heuristics were all embodied in the program. In this program the procedures will be described as much as possible in the language itself and, in particular, the heuristics are all so described.

The main advantages we expect the advice taker to have is that its behavior will be improvable merely by making statements to it, telling it about its symbolic environment and what is wanted from it. To make these statements will require little if any knowledge of the program or the previous knowledge of the advice taker. One will be able to assume that the advice taker will have available to it a fairly wide class of immediate logical consequences of anything it is told and its previous knowledge. This property is expected to have much in common with what makes us describe certain humans as having common sense. We shall therefore say that a program has common sense if it automatically deduces for itself a sufficiently wide class of immediate consequences of anything it is told and what it already knows.

The design of this system will be a joint project with Marvin Minsky, but Minsky is not to be held responsible for the views expressed here.

Before describing the advice taker in any detail, I would like to describe more fully our motivation for proceeding in this direction. Our ultimate objective is to make programs that learn from their experience as effectively as humans do. It may not be realized how far we are presently from this objective. It is not hard to make machines learn from experience to make simple changes in their behavior of a kind which has been anticipated by the programmer. For example, Samuel has included in his checker program facilities for improving the weights the machine assigns to various factors in evaluating positions. He has also included a scheme whereby the machine remembers games it has played previously and deviates from its previous play when it finds a position which it previously lost. Suppose, however, that we wanted an improvement in behavior corresponding, say, to the discovery by the machine of the principle of the opposition in checkers. No present or presently proposed schemes are capable of discovering phenomena as abstract as this.

If one wants a machine to be able to discover an abstraction, it seems most likely that the machine must be able to represent this abstraction in some relatively simple way.

There is one known way of making a machine capable of learning arbitrary behavior; thus to anticipate every kind of behavior. This is to make it possible for the machine to simulate arbitrary behaviors and try them out. These behaviors may be represented either by nerve nets (Minsky 1956), by Turing machines (McCarthy 1956), or by calculator programs (Friedberg 1958). The difficulty is two-fold. First, in any of these representations the density of interesting behaviors is incredibly low. Second, and even more important, small interesting changes in behavior expressed at a high level of abstraction do not have simple representations. It is as though the human genetic structure were represented by a set of blue-prints. Then a mutation would usually result in a wart or a failure of parts to meet, or even an ungrammatical blue-print which could not be translated into an animal at all. It is very difficult to see how the genetic representation scheme manages to be general enough to represent the great variety of animals observed and yet be such that so many interesting changes in the organism are represented by small genetic changes. The problem of how such a representation controls the development of a fertilized egg into a mature animal is even more difficult.

In our opinion, a system which is to evolve intelligence of human order should have at least the following features:

1. All behaviors must be representable in the system. Therefore, the system should either be able to construct arbitrary automata or to program in some general purpose programming language.
2. Interesting changes in behavior must be expressible in a simple way.
3. All aspects of behavior except the most routine must be improvable. In particular, the improving mechanism should be improvable.
4. The machine must have or evolve concepts of partial success because on difficult problems decisive successes or failures come too infrequently.
5. The system must be able to create subroutines which can be included in procedures as units. The learning of subroutines is complicated by the fact that the effect of a subroutine is not usually good or bad in itself. Therefore, the mechanism that selects subroutines should have concepts of interesting or powerful subroutine whose application may be good under suitable conditions.

Of the 5 points mentioned above, our work concentrates mainly on the second. We base ourselves on the idea that: In order for a program to be capable of learning something it must first be capable of being told it. In fact, in the early versions we shall concentrate entirely on this point and attempt to achieve a system which can be told to make a specific improvement in its behavior with no more knowledge of its internal structure or previous knowledge than is required in order to instruct a human. Once this is achieved, we may be able to tell the advice taker how to learn from experience.

The main distinction between the way one programs a computer and modifies the program and the way one instructs a human or will instruct the advice taker is this: A machine is instructed mainly in the form of a sequence of imperative sentences; while a human is instructed mainly in declarative sentences describing the situation in which action is required together with a few imperatives that say what is wanted. We shall list the advantages of of the two methods of instruction.

Advantages of Imperative Sentences

1. A procedure described in imperatives is already laid out and is carried out faster.
2. One starts with a machine in a basic state and does not assume previous knowledge on the part of the machine.

Advantages of Declarative Sentences

1. Advantage can be taken of previous knowledge.
2. Declarative sentences have logical consequences and it can be arranged that the machine will have available sufficiently simple logical consequences of what it is told and what it previously knew.
3. The meaning of declaratives is much less dependent on their order than is the case with imperatives. This makes it easier to have after-thoughts.
4. The effect of a declarative is less dependent on the previous state of the system so that less knowledge of this state is required on the part of the instructor.

The only way we know of expressing abstractions (such as the previous example of the opposition in checkers) is in language. That is why we have decided to program a system which reasons verbally.