Path: utzoo!attcan!uunet!yale!dvm
From: dvm@yale.UUCP (Drew Mcdermott)
Newsgroups: comp.ai
Subject: Free will
Keywords: philosophy,causality,free will
Message-ID: <30472@yale-celray.yale.UUCP>
Date: 30 May 88 16:40:28 GMT
Organization: Yale University, New Haven, CT
Lines: 245

More on the self-modeling theory of free will:

Since no one seems to have understood my position on this topic,
I will run the risk that no one cares about my position, and try
to clarify.

Sometimes parties to this discussion talk as if "free will" were
a new kind of force in nature.  (As when Biep Durieux proposed that
free will might explain probability rather than vice versa.)  I am
sure I misrepresent the position; the word "force" is surely wrong
here (as is the word "new").  The misrepresentation is unavoidable;
this kind of dualism is simply not a live option for me.  Nor can
I see why it needs to be a perenially live option on an AI discussion 
bulletin board.

So, as I suggested earlier, let's focus on the question of free will
within the framework of Artificial Intelligence.  And here it
seems to me the question is, How would we tell an agent with free
will from an agent without it?  Two major strands of the discussion
seem completely irrelevant from this standpoint:

  (1) Determinism vs. randomness.  The world is almost
certainly not deterministic, according to quantum mechanics.  Quantum
mechanics may be false, but Newtonian mechanics is certainly false,
so the evidence that the world is deterministic is negligible.
(Unless the Everett-Wheeler interpretation of quantum mechanics is true, 
in which case the world is a really bizarre place.)  So, if determinism
is all that's bothering you, you can relax.  Actually, I think what's
really bothering people is the possibility of knowledge (traditionally,
divine knowledge) of the outcomes of their future decisions, which has 
nothing to do with determinism.

  (2) My introspections about my ability to control my thoughts or 
whatnot.  There is no point in basing the discussion on such evidence,
until we have a theory of what conscious thoughts are.  Such a theory
must itself start from the outside, looking at a computational agent
in the world and explaining what it means for it to have conscious
thoughts.  That's a fascinating topic, but I think we can solve the
free will problem with less trouble.

  So, what makes a system free?  To the primitive mind, free decisions
are ubiquitous.  A tornado decides to blow my house down; it is worth
trying to influence its decision with various rewards or threats.
But nowadays we know that the concept of decision is just out of place
in reasoning about tornados.  The proper concepts are causal; if we
can identify enough relevant antecedent factors, we can predict (and
perhaps someday control) the tornado's actions.  Quantum mechanics
and chaos set limits to how finely we can predict, but that is
irrelevant.

  Now we turn to people.  Here it seems as if there is no need to do
away with the idea of decision, since people are surely the paradigmatic
deciders.  But perhaps that attitude is "unscientific."  Perhaps the
behaviorists are right, and the way we think about thunderstorms is
the right way to think about people.  If that's the actual truth, then
we should be tough-minded and acknowledge it.

  It is *not* the truth.  Freedom gets its toehold from the fact that
it is impossible for an agent to think of itself in terms of causality.  
Contrast my original bomb scenario with this one:  

   R sees C wander into the blast area, and go up to the bomb.  R knows
   that C knows all about bombs, and R knows that C has plenty of time to
   save itself, so R decides to do nothing.  (Assume that preventing the
   destruction of other robots gets big points in R's utility function.)

In this case, R is reasoning about an agent other than itself.  Its problem
is to deduce what C will actually do, and what C will actually suffer.  The
conclusion is that C will prosper, so R need do nothing.  It would
be completely inappropriate for R to reason this way about itself.  Suppose
R comes to realize that it is standing next to a bomb, and it reasons as
follows:

   R knows all about bombs, and has plenty of time to save itself, so I need
   do nothing.

Its reasoning is fallacious, because it is of the wrong kind.  R is not being
called on to deduce what R will do, but to be a part of the causal fabric that
determines what R will do, in other words: to make a decision.  It is certainly
possible for a robot to engage in a reasoning pattern of this faulty kind, but 
only by pretending to make a decision, inferring that the decision will be made 
like that, and then not carrying it out (and thus making the conclusion of the
inference false).  Of course, such a process is not that unusual; it is called
"weakness of the will" by philosophers.  But it is not the sort of thing one
would be tempted to call an actual decision.  An actual decision is a process
of comparative evaluation of alternatives, in a context
where the outcome of the comparison will actually govern behavior.  (A robot
cannot decide to stop falling off a cliff, and an alcoholic or compulsive may
not actually make decisions about whether to cease his self-destructive
behavior.)

   This scenario is one way for a robot to get causality wrong when reasoning
about itself, but there is a more fundamental way, and that is to just not 
notice that R is a decision maker at all.  With this misperception, R could 
tally its sources of knowledge about all influences on R's behavior, but it 
would miss the most important one, namely, the ongoing alternative-evaluation 
process. Of course, there are circumstances in which this process is in fact not
important.  If R is bound and gagged and floating down a river, then it might
as well meditate on hydrodynamics, and not work on a decision.  But most of the
time the decision-making process of the robot is actually one of the causal
antecedents of its future.  And hence, to repeat the central idea, *there is
no point in trying to think causally about oneself while making a decision that
is actually part of the causal chain.  Any system that realizes this has free
will.*

  This theory accounts for why an agent must think of itself as outside the
causal order of things when making a decision.  However, it need not think
of other agents this way.  An agent can perfectly well think of other agents' 
behavior as caused or uncaused to the same degree the behavior of a 
thunderstorm is caused or uncaused.  There is a difference: One of the best 
ways to cause a decision-making agent to do something is to give him a good 
reason to do it, whereas this strategy won't work with thunderstorms.  Hence, 
an agent will do well to sort other systems into two categories, those that 
make free decisions and those that don't, and deal with them differently.  

  By the way, once a decision is made there is no problem with its maker 
thinking of it purely causally, in exactly the same way it thinks about 
other decision makers.  An agent can in principle see *all* of the causal 
factors going into its own past decisions, although  in practice the events 
of the past will be too random or obscure for an exhaustive analysis.  It is 
surely not dehumanizing to be able to bemoan that if only such-and-such had 
been brought to my attention, I would have decided otherwise than I did, but, 
since it wasn't, I was led inexorably to a wrong decision.

 Now let me deal with various objections: 

   (1) Some people said I had neglected the ability of computers to do reflexive
meta-reasoning.  As usual, the mention of meta-reasoning makes my head swim, but
I shall try to respond.  Meta-reasoning can mean almost anything, but it usually
means escaping from some confining deductive system in order to reason about 
what that system ought to conclude.  If this is valuable, there is no reason 
not to use it.  But my picture is of a robot faced with the possibility of 
reasoning about itself as a physical system, which is in general a bad thing to 
do.  The purpose of causal-exemption flagging is to shut pointless reasoning 
down, meta or otherwise.

So, when O'Keefe says:

    So the mere possibility of an agent having to appear to simulate itself 
    simulating itself ... doesn't show that unbounded resources would be 
    required:  we need to know more about the nature of the model and the 
    simulation process to show that.

I am at a loss.  Any system can simulate itself with no trouble.  It could 
go over past or future decisions with a fine-tooth comb, if it wanted to.  
What's pointless is trying to simulate the present period of time.  Is an 
argument needed here?  Draw a mental picture: The robot starts to simulate, 
and finds itself simulating ...  the start of a simulation.  What on earth 
could it mean for a system to figure out what it's doing by simulating itself?

  (2) Free will seems on this theory to have little to do with consciousness 
or values.  Indeed it does not.  I think a system could be free and not be 
conscious at all; and it could certainly be free and not be moral.

  What is the minimal level of free will?  Consider a system for scheduling the
movement of goods into and out of a warehouse.  It has to synchronize its 
shipments with those of other agents, and let us suppose that it is given 
those other shipments in the form of various schedules that it must just work 
around.  From its point of view, the shipments of other agents are caused, and 
its own shipments are to be selected.  Such a system has what we might call 
*rudimentary* free will.  To get full-blown free will, we have to suppose that 
the system is able to notice the discrepancy between boxes that are scheduled 
to be moved by someone else, and boxes whose movements depend on its 
decisions.  I can imagine all sorts of levels of sophistication in 
understanding (or misunderstanding) the discrepancy, but just noticing it is 
sufficient for a system to have full-blown free will.  At that point, it will 
have to realize that it and its tools (the things it moves in the warehouse) 
are exempt from causal modeling.

   (3) Andrew Gelsey has pointed out that a system might decide what to do by
means other than simulating various alternative courses of action.  For 
instance, a robot might decide how hard to hit a billiard ball by solving an 
equation for the force required.  In this case, the asymmetry appears in what 
is counted as an independent variable (i.e., the force administered).  And if 
the robot notices and appreciates the asymmetry, it is free.

   (4) David Sher has objected

      If I understand [McDermott's theory] correctly it runs like this:
      To plan one has a world model including future events.
      Since you are an element of the world then you must be in the model.
      Since the model is a model of future events then your future actions
      are in the model.
      This renders planning unnecessary.
      Thus your own actions must be excised from the model for planning to
      avoid this "singularity."

      Taken naively, this analysis would prohibit multilevel analyses such
      as is common in game theory.  A chess player could not say things like
      if he moves a6 then I will move Nc4 or Bd5 which will lead ....

The response to this misreading should be obvious.  There are two ways to
think about my future actions.  One way is to treat them as conditional actions,
begun now, and not really future actions at all.  (Cf. the notion of strategy
in game theory.)

The more interesting insight is that an agent can reason about its future 
actions as if they were those of another agent.  There is no problem with 
doing this; the future is much like the past in this respect, except we have 
less information about it.  A robot could reason at its leisure about what 
decision it would probably make if confronted with some future situation, and 
it could use an arbitrarily detailed simulation of itself to do this reasoning, 
provided it has time to run it before the decision is to be made.  But all of 
this self-prediction is independent of actually making the decision.  When 
the time comes to actually make it, the robot will find itself free again.  
It will not be bound by the results of its simulation.  This may seem like a 
nonsequitur; how could a robot not faithfully execute its program the same 
way each time it is run?  There is no need to invoke randomness; the 
difference between the two runs is that the second one is in a context where 
the results of the simulation are available.  Of course, there are lots of 
situations where the decision would be made the same way both times, but all 
we require is that the second be correctly classified as a real -- free -- 
decision.

I find Sher's "fix" to my theory more dismaying:

   However we can still make the argument that Drew was making its just
   more subtle than the naive analysis indicates.  The way the argument 
   runs is this:
   Our world model is by its very nature a simplification of the real
   world (the real world doesn't fit in our heads).  Thus our world model
   makes imperfect predictions about the future and about consequence.  
   Our self model inside our world model shares in this imperfection. 
   Thus our self model makes inaccurate predictions about our reactions
   to events.  We perceive ourselves as having free will when our self
   model makes a wrong prediction.  

This is not at all what I meant, and seems pretty shaky on its own merits.
This theory makes an arbitrary distinction between an agent's mistaken 
predictions about itself and its mistaken predictions about other systems.
I think it's actually a theory of why we tend to attribute free will to so
many systems, including thunderstorms.  We know our freedom makes us hard to
predict, and so we attribute freedom to any system we make a wrong prediction
about.  This kind of paranoia is probably healthy until proven false.  But
the theory doesn't explain what we think free will is in the first place, or
what its explanatory force is in explaining wrong predictions.

Free will is not due to ignorance.  Imagine that the decision maker is a robot
with a very routine environment, so that it often has complete knowledge both of
its own listing and of the external sensory data it will be receiving prior to
a decision.  So it can simulate itself to any level of detail, and it might
actually do that, thinking about decisions in advance as a way of saving time 
later when the actual decision had to be made.  None of this would allow it to 
avoid making free decisions.  
                                     -- Drew McDermott