Programs with Common Sense
My first real Blog post will be about John McCarthy's classic paper Programs with Common Sense which may be regarded as the canonical description of a Good Old Fashioned Artificial Intelligence. John McCarthy, of course, was one of the original founders of our discipline, and came up with the name "Artificial Intelligence" (He's also my advisor's advisor, my grand-advisor if you will.)
In this paper, McCarthy lays out his vision for a thinking machine he calls the Advice Taker. When he and Marvin Minsky started working on this project, there was already talk of constructing intelligent machines, most notably the Logic Theory Machine of Newell, Simon and Shaw. All these proposals involved a reasoning system that used heuristics that were preprogrammed. McCarthy envisioned a machine whose heuristics would be in the representation language itself. Thus it's Behaviour could be improved merely by gaining knowledge about the outside world. On this basis McCarthy defined Common Sense - the ability of the machine to automatically deduce for itself a sufficiently wide class of immediate consequences of anything it is told and what it already knows.
To do so, he realised, one needs a language where interesting changes in behaviour could be expressible in a simple way. Thus he argued, the use of declarative sentences to express knowledge should be favored over imperative sentences to guide action. This was the same rationale that led to the invention of Lisp.
He goes on to describe the function of the Advice Taker in detail. It would use the First Order Predicate Calculus, the representation that dominated Knowledge Representation in AI research until today. Even by the standards of the next decade, the inference system was ridiculously naive. Modus Ponens was its one inference rule.
But as he demonstrated with his want(at(I,airport)) example, the deductive power of such a system was potentially immense. He also pointed out the fundamental stumbling block that would plague GofAI - the inherent non-determinism of reasoning. McCarthy realised that the Common-Sense of the machine would be a function of how good its heuristics would be in picking out the "right" line of deduction.
The discussion after the presentation of the paper is interesting as well. It seems that Professor Y. Bar-Hillel who dismissed Mccarthy's ideas as "half-baked", misunderstood his work and did not truly appreciate the scope of what McCarthy was attempting. However a couple of his criticisms were right on the mark and foreshadowed the identification of the qualification problem and the frame problem, which were dealt with in more detail in Mccarthy and Hayes 1969 paper Some Philosophical Problems From the Standpoint of Artificial Intelligence, as a primary weakness of the system.
Looking back now, the idea of a Software program with a large knowledge bank of rules, chugging along grinding out conclusions and dispensing advice, seems way too simplistic. But the real contribution of this paper was giving a goal and a direction to AI research that stimulated computer scientists for the next couple of decades.
In this paper, McCarthy lays out his vision for a thinking machine he calls the Advice Taker. When he and Marvin Minsky started working on this project, there was already talk of constructing intelligent machines, most notably the Logic Theory Machine of Newell, Simon and Shaw. All these proposals involved a reasoning system that used heuristics that were preprogrammed. McCarthy envisioned a machine whose heuristics would be in the representation language itself. Thus it's Behaviour could be improved merely by gaining knowledge about the outside world. On this basis McCarthy defined Common Sense - the ability of the machine to automatically deduce for itself a sufficiently wide class of immediate consequences of anything it is told and what it already knows.
To do so, he realised, one needs a language where interesting changes in behaviour could be expressible in a simple way. Thus he argued, the use of declarative sentences to express knowledge should be favored over imperative sentences to guide action. This was the same rationale that led to the invention of Lisp.
He goes on to describe the function of the Advice Taker in detail. It would use the First Order Predicate Calculus, the representation that dominated Knowledge Representation in AI research until today. Even by the standards of the next decade, the inference system was ridiculously naive. Modus Ponens was its one inference rule.
But as he demonstrated with his want(at(I,airport)) example, the deductive power of such a system was potentially immense. He also pointed out the fundamental stumbling block that would plague GofAI - the inherent non-determinism of reasoning. McCarthy realised that the Common-Sense of the machine would be a function of how good its heuristics would be in picking out the "right" line of deduction.
The discussion after the presentation of the paper is interesting as well. It seems that Professor Y. Bar-Hillel who dismissed Mccarthy's ideas as "half-baked", misunderstood his work and did not truly appreciate the scope of what McCarthy was attempting. However a couple of his criticisms were right on the mark and foreshadowed the identification of the qualification problem and the frame problem, which were dealt with in more detail in Mccarthy and Hayes 1969 paper Some Philosophical Problems From the Standpoint of Artificial Intelligence, as a primary weakness of the system.
Looking back now, the idea of a Software program with a large knowledge bank of rules, chugging along grinding out conclusions and dispensing advice, seems way too simplistic. But the real contribution of this paper was giving a goal and a direction to AI research that stimulated computer scientists for the next couple of decades.
posted by Deepak at 3:45 PM
0 Comments:
Post a Comment
Links to this post:
Create a Link
<< Home