Date: Mon 1 Aug 1988 14:37-EDT From: AIList Moderator Nick Papadakis Reply-To: AIList@mc.lcs.mit.edu Us-Mail: MIT Mail Stop 38-390, Cambridge MA 02139 Phone: (617) 253-2737 Subject: AIList Digest V8 #34 To: AIList@mc.lcs.mit.edu Status: R AIList Digest Tuesday, 2 Aug 1988 Volume 8 : Issue 34 Today's Topics: Free Will: How to dispose of naive science types (short) The deterministic robot determines that it needs to become nondeterministic. Root issue of free will and problems in war zones ---------------------------------------------------------------------- Date: 27 Jul 88 09:09:44 GMT From: mcvax!ukc!strath-cs!glasgow!gilbert@uunet.uu.net (Gilbert Cockton) Subject: Re: How to dispose of naive science types (short) In article <531@ns.UUCP> logajan@ns.UUCP (John Logajan x3118) writes: >Please explain to me how an unproveable theory (one that makes no unique >predictions) can be useful? > Because people use them. Have a look at the social cognition literature. I understood your argument as saying that non-scientific theories (a.k.a assumptions) cannot be useful, and conversely, that the only useful theories are scientific ones. If my understanding is correct, then this is very narrow minded and smacks of epistemelogical bigotry which no-one can possibly match up to in their day to day interactions. Utility must not be confounded with one text-book epistemology. -- Gilbert Cockton, Department of Computing Science, The University, Glasgow gilbert@uk.ac.glasgow.cs !ukc!glasgow!gilbert ------------------------------ Date: 27 Jul 88 15:34:09 GMT From: bwk@mitre-bedford.ARPA (Barry W. Kort) Reply-to: bwk@mbunix (Kort) Subject: The deterministic robot determines that it needs to become nondeterministic. In article <19880727030413.0.NICK@HOWARD-JOHNSONS.LCS.MIT.EDU> JMC@SAIL.STANFORD.EDU (John McCarthy) writes: >Almost all the discussion is too vague to be a contribution. Let me >suggest that AI people concentrate their attention on the question of how >a deterministic robot should be programmed to reason about its own free >will, as this free will relates both to its past choices and to its future >choices. Can we program it to do better in the future than it did in the >past by reasoning that it could have done something different from what it >did, and this would have had a better outcome? If yes, how should it be >programmed? If no, then doesn't this make robots permanently inferior to >humans in learning from experience? To my mind, the robot's problem becomes interesting precisely when it runs out of knowledge to predict the outcome of the choices open to it. The classical metaphors for this state are "The Lady or the Tiger?", the Parable of Buridan's Ass, and "Dorothy meets the Scarecrow at the fork in the road." The children's game of Rock, Scissors, Paper illustrates the predicament faced by a deterministic robot. In the above scenarios, the resolution is to pick a path at random and pursue it first. To operationalize the decision, one needs to implement the Axiom of Choice. One needs a random number generator. Fortunately, it is possible to build one using a Quantum Amplifier. (Casting lots will do, if you live in a low-tech society.) Thus, I conclude that a deterministic robot will perceive itself at a disadvantage relative to a robot who can implement the Axiom of Choice, and will decide (of its own free will) that it must evolve to include nondeterministic behavior. Note, by the way, that decision in the face of uncertainty entails a risk, so a byproduct of such behavior is anxiety. In other words, emotion is the expression of vanishing ignorance. --Barry Kort ------------------------------ Date: Wed 27 Jul 88 13:53:59-PDT From: Leslie DeGroff Subject: root issue of free will and problems in war zones The new AI in the war zone and the on going free will discussions seem to both skirt around one of the fundamental crux's of Intelligence, natural and artificial (and even "Non Intelligent" decision making processes) There is a pair of Quantities that appear in all decision processes, one is the information/knowledge in the "system/agent/individual" and the other is the scale and variation of the universe to be modeled. For the real world the latter is always much much greater than the prior. Universe >> some subsystem. Even if we take out infinities there is this many order of magnitude scale problem. This inequally holds regaurdless of the equivalence of the internal representation to the external "facts" Engineers, Programmers, and line managers get their noses rubbed in this fact pretty often (but perhaps not enough to prevent horrible/scary/dumb mistakes from being made) This ratio more or less means that systems working in the real world can always be surprised and/or make mistakes. The universe does have regularities that allow the causal and structural mapping of a smaller "Mind" or "representation" to cover a lot of ground but it also remains filled with places where you need to know the specifics to know what is happening. Even simple Newtonian physics of multiple orbiting bodies becomes a combinatorial problem very quickly. In regards to the war zone, we have a similar case (the Russians and KAL) which had no particular computer component... Just miss or missing communications/information and a human decision. There is a limit to the precision and availability of knowledge and an even lower limit to the amount of processing that can be done. The universe and Murphy will get us everytime we start thinking "it's ALL under control". Related to this fundamental fact is that in many cases "WILL" turns out to be a concept used by humans to represent the immediate uncomputability/unpredictability of peices of the real Universe including our own actions and conciousness. I find WILL to be a much more productive concept to contemplate than FREE WILL. I can be scientifically educated and still talk and think of inanimate objects like a truck or a storm as having willful behavior. Even simple physical systems with unsensed or unpredictable variability will often be treated as if decisions are being made; ?Will my door handle give me a static spark today? Much of the discussion on determinism vs non is simply missing the point that neither our brains nor our computers will be able to "compute" in real time all that might be of importance to a given situation and no realistic set of sensors can gather enough information to be "complete". From an AI perspective these issues are at the heart of the hardness in the problems; how can we have an open ended learning system with out catatonic behavior? (computation of all derivations from an ever increasing fact base)and what kind of knowledge representation is efficient for learning from sensors, effective at cutting off computation so that time critical decisions can be made and knowing when knowledge contained dosn't apply (classic case of the potential infinity of negations) (Trick question for the brain modelers, Does sleep act like a Lisp Garbage collector ie is part of the sleep process an elimination of material that is not to be stored and reorganizing the rest of the material) Much of applied statistics and measurement theory is oriented to METRICs for comparing systems and models and determining "predicts correctly and fails to predict" where the models are parametric equations. Question is how to evaluate a model for "surprise potential" or "unincluded critcal factors". Les Degroff DeGroff@intellicorp.com (I disclaim all blame, I aint paid to think but I have this bad habit, neither parents, schools or employers have been able to cure it) ------------------------------ Date: 28 Jul 88 18:20:16 GMT From: umix!umich!eecs.umich.edu!itivax!dhw@uunet.UU.NET (David H. West) Subject: Re: free will In a previous article, John McCarthy writes: > Let me > suggest that AI people concentrate their attention on the question of how > a deterministic robot should be programmed to reason about its own free > will, as this free will relates both to its past choices and to its future > choices. Can we program it to do better in the future than it did in the > past by reasoning that it could have done something different from what it > did, and this would have had a better outcome? If yes, how should it be > programmed? If no, then doesn't this make robots permanently inferior to > humans in learning from experience? At time t0, the robot has available a partial (fallible) account of: the world-state, its own possible actions, the predicted effects of these actions, and the utility of these effects. Suppose it wants to choose the action with maximum estimated utility, and further suppose that it can and does do this. Then its decision is determined. Free-will (whatever that is) is merely the freedom to do something that doesn't maximize its utility, which is ex hypothesi not a freedom worth exercising. At a later time t1, the robot has available all of the above, plus the outcome of its action. It is therefore not in the same state as previously. It would make no sense to ignore the additional information. If the outcome was as expected, then there is no reason to make a different choice next time unless some other element of the situation changes. If the outcome was not as predicted, the robot needs to update its models. This updating is another choice-of-action-under-incomplete-information problem, so again the robot can only maximize its own fallibly-estimated utility, and again its behavior is determined, not (just) by its physical structure, but by the meta-goal of acting coherently. If the robot thought about its situation, it would presumably conclude that it felt no impediment to doing what was obviously the correct thing to do, and that it therefore had free will. -David West dhw%iti@umix.cc.umich.edu ------------------------------ Date: 29 Jul 88 18:18:48 GMT From: well!sierch@lll-lcc.llnl.gov (Michael Sierchio) Subject: Re: How to dispose of naive science types (short) Theories are not for proving! A theory is a model, a description, an attempt to preserve and describe phenomena -- science is not concerned with "proving" or "disproving" theories. Proof may have a slightly different meaning for attorneys than for mathematicians, but scientists are closer to the mathematician's definition -- when they use the word at all. A theory may or may not adequately describe the phenomena in question, in which case it is a "good" or "bad" theory -- of two "good" theories, the theory that is "more elegant" or "simpler" may be preferred -- but this is an aesthetic or performance judgement, and again has nothing to do with proof. Demonstration and experimentation show (to one degree or another) the value of a particular theory in a particular domain -- but PROOF? bah! -- Michael Sierchio @ Small Systems Solutions sierch@well.UUCP {pacbell,hplabs,ucbvax,hoptoad}!well!sierch ------------------------------ End of AIList Digest ********************