Path: utzoo!utgpu!water!watmath!clyde!att!osu-cis!tut.cis.ohio-state.edu!bloom-beacon!AI.AI.MIT.EDU!NICK From: NICK@AI.AI.MIT.EDU (Nick Papadakis) Newsgroups: comp.ai.digest Subject: [JMC@SAIL.Stanford.EDU: free will ] Message-ID: <19880528031924.2.NICK@MACH.AI.MIT.EDU> Date: 28 May 88 03:19:00 GMT Sender: daemon@bloom-beacon.MIT.EDU Organization: The Internet Lines: 67 Approved: ailist@ai.ai.mit.edu Date: Wed, 25 May 88 10:32 EDT From: John McCarthy Subject: free will To: ailist@AI.AI.MIT.EDU The following propositions are elaborated in {\bf McCarthy, John and P.J. Hayes (1969)}: ``Some Philosophical Problems from the Standpoint of Artificial Intelligence'', in D. Michie (ed), {\it Machine Intelligence 4}, American Elsevier, New York, NY. I would be grateful for discussion of them - especially technical discussion. 1. For AI, the key question concerning free will is "What view should we program a robot to have of its own free will?". I believe my proposal for this also sheds light on what view we humans should take of our own free will. 2. We have a problem, because if we put the wrong assertions in our database of common sense knowledge, a logic-based robot without a random element might conclude that since it is a deterministic robot, it doesn't make sense for it to consider alternatives. It might reason: "Since I'm a robot, what I will do is absolutely determined, so any consideration of whether one course of action or another would violate (for example) Asimov's suggestion that robots shouldn't harm human beings is pointless". 3. Actually (McCarthy and Hayes 1969) considered an even more deterministic system than a robot in the world - namely a system of interconnected finite automata and asked the question: "When should we say that in a given initial situation, automaton 1 can put automaton 7 in state 3 by time 10?" 4. The proposed answer makes this a definite question about another automaton system, namely a system in which automaton 1 is removed from the original system, and its output lines are replaced by external inputs to the revised system. We then say that automaton 1 can put automaton 7 in state 3 by time 10 provided there is a sequence of signals on the external inputs to the revised system that will do it. 5. I claim this is how we want the robot to reason. We should program it to decide what it can do, i.e. the variety of results it can achieve, by reasoning that doesn't involve its internal structure but only its place in the world. Its program should then decide what to do based on what will best achieve the goals we have also put in its database. 6. I claim that my own reasoning about what I can do proceeds similarly. I model the world as a system of interacting parts of which I am one. However, when deciding what to do, I use a model in which my outputs are external inputs to the system. 7. This model says that I am free to do those things that suitable outputs will do in the revised system. I recommend that any "impressionable students" in the audience take the same view of their own free will. In fact, I'll claim they already do; unless mistaken philosophical considerations have given them theories inferior to the most naive common sense. 8. The above treats "physical ability". An elaboration involving knowledge, i.e. that distinguishes my physical ability to dial your phone number from my epistemological ability that requires knowing the number, is discussed in the paper. These views are compatible with Dennett's and maybe Minsky's. In my view, McDermott's discussion would be simplified if he incorporated discussion of the revised automaton system.