Cognitive Science Goes Off the Tracks (was Re: PCT in context)

[From Rick Marken (2010.03.20.1100)]

Bruce Gregory (2010.03.19.1734 EDT)--

... Much of what today is called cognitive science and cognitive
neuroscience lies outside the domain explored by PCT.

I would like to explore this a bit further. While doing research for a
talk that I presented at the UCLA Psychology Department's Cognitive
Forum (in January, 2008) I discovered (or, perhaps, rediscovered)
something interesting about the development of what is now called
"cognitive science". One of the main streams feeding the development
of this field was the computer modeling of cognitive processes that
was being pioneered (in the early 1960s at RAND Corp.) by Newell and
Simon with their chess playing and General Problem Solving (GPS)
programs. What is interesting about these programs (and the
"artificial intelligence" programs that followed) is that they are
architected as hierarchical input control systems. The programs have
high level goals (like "control of center") that set subgoals
("advance knights") and then make moves to achieve the subgoals and,
as a result, the goals. So the first cognitive science models were
control models (though they didn't think of them that way).

Cognitive science went off the tracks when it moved from testing their
theories by seeing if they could build successful computer simulations
(which were organized as control systems) to testing their theories
using conventional experimental methods. One piece of evidence for
this is a study done by Chase and Simon (1973, well after Simon's GPS
simulation work) where they studied the ability of chess masters and
novices to remember chess positions. What they found is that chess
masters are better (on average; this was a multi-subject experiment,
of course) than novices at remembering real chess positions but that
both are equal at remembering positions formed by randomly placing
pieces on the board. Their conclusion was that chess skill results
from remembering chess positions and the appropriate "answering" moves
and that masters are better than novices because the perception of a
chess position evokes more useful remembered positions for the master
than for the novices.

So here we have a researcher who had started by (properly, I believe)
conceiving of chess playing (and problem solving in general) as a
purposeful behavior, where the problem is solved by acting (making
moves) to bring the state of the problem closer to subgoals and goal
states. And then, after doing research in an S-R framework (as is all
conventional research) comes up with this S-R theory of problem
solving: with the perception of board position (S) evoking a memory
associated with the answering move.

As I read the Chase/Simon paper I could almost hear the crash of the
cognitive train going off the tracks. If only they had understood how
to study control they would have been able to test their control
models appropriately and not gotten derailed by S-R thinking. For
example, if the chess playing model has a goal of "controlling for the
center" then presumably this goal is defined quantitatively in the
program. Then it should be pretty straight forward to design some
tests -- disturbing this hypothetical controlled variable by making
moves that increase or decrease control of the center -- to see if a
chess player is, indeed, controlling for this variable.

Now that we understand how control theory applies to behavior
(including "cognitive" behavior like problem solving) maybe we can get
the cognitive science train back on track. But it will take a lot more
work (and a lot less talk).

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.20.1430 EDT)]

[From Rick Marken (2010.03.20.1100)]

So here we have a researcher who had started by (properly, I believe)
conceiving of chess playing (and problem solving in general) as a
purposeful behavior, where the problem is solved by acting (making
moves) to bring the state of the problem closer to subgoals and goal
states. And then, after doing research in an S-R framework (as is all
conventional research) comes up with this S-R theory of problem
solving: with the perception of board position (S) evoking a memory
associated with the answering move.

One of the nice things about a control system is that it does not need much memory. All it needs to remember are reference levels for controlled variables. The thermostat in my heating system has no memory and it works perfectly. Too bad so much of the brain is devoted to memory when it could be devoted to something useful. Maybe the next step in evolution will be a more efficient design.

Bruce

[From Bill Powers (2010.03.201725 MDT)]

From Bruce Gregory (2010.03.20.1430 EDT) --

BG: One of the nice things about a control system is that it does not need much memory. All it needs to remember are reference levels for controlled variables. The thermostat in my heating system has no memory and it works perfectly. Too bad so much of the brain is devoted to memory when it could be devoted to something useful. Maybe the next step in evolution will be a more efficient design.

BP: Bruce, I challenge you to find anything I or Rick have written which says that memory is used only for reference signals. You won't find it.

The reason for involving memory with reference signals is very simple and obvious. I have explained it many times, in things I'm sure you have read (like B:CP, and in discussions on CSGnet). If you just pause and think about it yourself, I'm sure you are capable of seeing why and when reference signals come from memory. This doesn't mean that all reference signals come from memory, or that the only function of memory is to provide reference signals. It just means that in certain cases, reference signals obviously come from memory. We don't know, of course, how extensive this use of memory is, or whether it happens at all levels, or as I have suggested previously, primarily at the higher levels and probably not at all at the lowest levels.

Bill P.

[From Rick Marken (2010.03.20.1730)]

Bruce Gregory (2010.03.20.1430 EDT) re: Rick Marken (2010.03.20.1100)--

One of the nice things about a control system is that it does not need much
memory...

You'll be happy to know that most of the folks who attended my
Cognitive Forum presentation -- distinguished cognitive scientists and
neuroscientists and their grad students -- didn't like what I said any
better than you did. So you must be on the right track;-)

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.20.2054 EDT)]

[From Bill Powers (2010.03.201725 MDT)]

From Bruce Gregory (2010.03.20.1430 EDT) --

BG: One of the nice things about a control system is that it does not need much memory. All it needs to remember are reference levels for controlled variables. The thermostat in my heating system has no memory and it works perfectly. Too bad so much of the brain is devoted to memory when it could be devoted to something useful. Maybe the next step in evolution will be a more efficient design.

BP: Bruce, I challenge you to find anything I or Rick have written which says that memory is used only for reference signals. You won't find it.

BG: I didn't say you did. I said that the only thing a hierarchical control system needs to remember are reference signals. Is that incorrect? What input from memory is required for the operation of a hierarchical control system other than a reference signal? I don't see it on any diagram. What am I missing?

The reason for involving memory with reference signals is very simple and obvious. I have explained it many times, in things I'm sure you have read (like B:CP, and in discussions on CSGnet). If you just pause and think about it yourself, I'm sure you are capable of seeing why and when reference signals come from memory.

BG: That's what I said. Or so I thought.

This doesn't mean that all reference signals come from memory,

BG: Of course not. Some reference signals are the result of reorganization.

or that the only function of memory is to provide reference signals. It just means that in certain cases, reference signals obviously come from memory. We don't know, of course, how extensive this use of memory is, or whether it happens at all levels, or as I have suggested previously, primarily at the higher levels and probably not at all at the lowest levels.

BG: I still don't see what you are objecting to. Memory may serve other functions, but those functions do not play a role in the operation of a hierarchical control system. Or have I, once again, missed something that should be obvious?

Bruce

[From Bruce Gregory (2010.03.20.2056)]

[From Rick Marken (2010.03.20.1730)]

Bruce Gregory (2010.03.20.1430 EDT) re: Rick Marken (2010.03.20.1100)–

One of the nice things about a control system is that it does not need much
memory…

You’ll be happy to know that most of the folks who attended my
Cognitive Forum presentation – distinguished cognitive scientists and
neuroscientists and their grad students – didn’t like what I said any
better than you did. So you must be on the right track;-)

They were probably upset to realize that chess masters are wasting their time replaying the games of other masters when they could be doing something useful.

Bruce

[From Rick Marken (2010.03.20.1800)]

Bill Powers (2010.03.201725 MDT)\--

Bruce Gregory (2010.03.20.1430 EDT) --

BG: One of the nice things about a control system is that it does not need
much memory...

BP: Bruce, I challenge you to find anything I or Rick have written which
says that memory is used only for reference signals. You won't find it.

I don't think Bruce really even reads what we say. He's more like a
Fox News interviewer, picking out irrelevant phrases for political
spin purposes. My point about the Chase and Simon (1973) paper was
obviously not about memory; it was about their conclusion that move
selection in chess was an S-R process (as it seems to be when one
looks at the results of their conventional experiment) rather than a
goal oriented process (as it is implemented in Simon's own chess
playing programs). I based my comment on something BG had said but the
post was not really for him; I just thought it might be an interesting
observation to share on the net.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.20.2115 EDT)]

[From Rick Marken (2010.03.20.1800)]

I don’t think Bruce really even reads what we say. He’s more like a
Fox News interviewer, picking out irrelevant phrases for political
spin purposes.

BG: Flattery will get you nowhere.

My point about the Chase and Simon (1973) paper was
obviously not about memory; it was about their conclusion that move
selection in chess was an S-R process (as it seems to be when one
looks at the results of their conventional experiment) rather than a
goal oriented process (as it is implemented in Simon’s own chess
playing programs). I based my comment on something BG had said but the
post was not really for him; I just thought it might be an interesting
observation to share on the net.

BG: You are one who sees S-R processes everywhere. What Chase and Simon were arguing is that chess masters can remember thousands of positions and the moves that were made in the context of those positions. There is no need to assume that because they can do this, chess masters are S-R machines. In fact, I think you will find that what they remember allows them to drastically prune the number of options they explore before making a move. Hierarchical control systems have no need for extensive memories. But if I am wrong about this, perhaps you can point me to the place where such a need is reflected in a working model.

Bruce

[From Rick Marken (2010.03.21.1400)]

Bruce Gregory (2010.03.20.2115 EDT)--

Rick Marken (2010.03.20.1800)]

My point about the Chase and Simon (1973) paper was
obviously not about memory; it was about their conclusion that move
selection in chess was an S-R process

BG: You are one who sees S-R processes everywhere. What Chase and Simon were
arguing is that chess masters can remember thousands of positions and the
moves that were made in the context of those positions. There is no need to
assume that because they can do this, chess masters are S-R machines. In
fact, I think you will find that what they remember allows them to
drastically prune the number of options they explore before making a move.

If what you are saying is that a person with many position-move
associations stored in memory will have to do more drastic pruning
than a person with fewer such associations in memory in order to get
down to one position-move association then I certainly agree. But
pruning (eliminating options) implies a control process, one with the
goal of finding the one position-move association (out of the many in
memory) that is "right" (that matches a reference for what the move
should be at that point in the game). So you seem to imply that the
Chase/Simon model of chess playing is a control mode (like Simon's
computer simulation). If this is so then what their experiment reveals
is rather trivial: master players select the desired (reference) move
by sorting through more options than do novices. The experiment
doesn't tell us much about the control process that make it possible
for people (novices or masters) to play chess.

Hierarchical control systems have no need for extensive memories.

They need it only to do things that require extensive memories, like
Bill's "control for the president" experiment that he mentioned
earlier. In order to carry out that task a person has to have in
memory the names of all US presidents in the order in which they occur
(making controlling for the guy between Grover Cleveland's two terms a
bit tough;-)

But if I am wrong about this, perhaps you can point me to the place where such a need
is reflected in a working model.

A model of the "president" control task described above would need a
pretty extensive memory.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.21.1752 EDT)]

[From Rick Marken (2010.03.21.1400)]

Bruce Gregory (2010.03.20.2115 EDT)–

BG: You are one who sees S-R processes everywhere. What Chase and Simon were
arguing is that chess masters can remember thousands of positions and the
moves that were made in the context of those positions. There is no need to
assume that because they can do this, chess masters are S-R machines. In
fact, I think you will find that what they remember allows them to
drastically prune the number of options they explore before making a move.

If what you are saying is that a person with many position-move
associations stored in memory will have to do more drastic pruning
than a person with fewer such associations in memory in order to get
down to one position-move association then I certainly agree.

BG: That was not my point, but I’m glad you agree. The evidence is that chess players rely on pattern recognition to prune the possibilities. Pattern recognition can be described as a closed-loop process, needless to say.

But
pruning (eliminating options) implies a control process, one with the
goal of finding the one position-move association (out of the many in
memory) that is “right” (that matches a reference for what the move
should be at that point in the game).

BG: So the player recognizes the correct move because it matches a reference for the correct move at that point in the game. Devilishly clever I’d say. I think I now understand. Every memory is a reference for a control loop whose function is to identify that particular memory. The next step is obvious, every S-R process can be interpreted as a closed-loop process with a reference level set to R. I have no reason to question that claim.

Bruce

[From Rick Marken (2010.03.22.0920)]

Bruce Gregory (2010.03.21.1752 EDT)--

Rick Marken (2010.03.21.1400)--

BG: That was not my point, but I'm glad you agree. The evidence is that
chess players rely on pattern recognition to prune the possibilities.
Pattern recognition can be described as a closed-loop process, needless to
say.

Could you explain how pattern recognition prunes the (position-move, I
presume) possibilities?

But pruning (eliminating options) implies a control process, one with the
goal of finding the one position-move association (out of the many in
memory) that is "right" (that matches a reference for what the move
should be at that point in the game).

BG: So the player recognizes the correct move because it matches a reference
for the correct move at that point in the game.

No, the reference is the desired (and, thus, "right" from the point of
view of the pruner) position-move. It's what makes the pruning
possible. Without a specification of the position-move desired how
could you know which ones to remove (prune). It's like pruning a tree;
you have to know which branches to keep in order to know which ones to
remove.

But maybe the pattern-recognition approach can prune position-moves
without knowing which ones to keep. It would like to find out how that
works.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bill Powers (2010.03.22.1111 MDT)]

Rick Marken (2010.03.22.0920) --

No, the reference is the desired (and, thus, "right" from the point of
view of the pruner) position-move. It's what makes the pruning
possible. Without a specification of the position-move desired how
could you know which ones to remove (prune). It's like pruning a tree;
you have to know which branches to keep in order to know which ones to
remove.

This isn't true if the pruning results from reorganization. Look at Demo 8-1, ArmControlReorg. In the bottom right quadrant there's an "Output Weight Matrix" showing "From system" numbers down the left side, and "To environmental variable" numbers across the top. When you start the program running with reorganization turned on, this 14 x 14 matrix shows up as little rectangles of different brightnesses. The brightnesses show the relative weighting of the connection from each control system's output to each environmental variable. After about 30 seconds, almost all of the rectangles will have some brightness greater than the initial black. This means that each system is affecting all 14 joint angles of the arm (the "environmental variables"). The arm movements are, of course, quite random at this point.

The rest of the reorganizing process amounts to pruning out the unwanted connections and increasing the weights of the ones needed to control each joint angle. This is done by E. coli reorganization with each control system's error signal being the intrinsic variable being controlled, and control being achieved by variations of the output weights in the path to each environmental variable. When reorganization is essentially complete (in 20 minutes or half an hour), only the diagonal rectangles are lit up, showing that each control system is affecting only the joint angle that it's sensing.

So pruning doesn't have to be a systematic conscious process.

Best,

Bill P.

[From Rick Marken (2010.03.22.1120)]

Bill Powers (2010.03.22.1111 MDT)--

Rick Marken (2010.03.22.0920) --

No, the reference is the desired (and, thus, "right" from the point of
view of the pruner) position-move. It's what makes the pruning
possible. Without a specification of the position-move desired how
could you know which ones to remove (prune). It's like pruning a tree;
you have to know which branches to keep in order to know which ones to
remove.

This isn't true if the pruning results from reorganization.

Yes. I was assuming that this pruning was being done by a skilled chess player.

So pruning doesn't have to be a systematic conscious process.

Yes, the process I was thinking of was systematic (not necessarily conscious).

The pruning that is done by reorganization is random (unsystematic)
but it is still part of a closed loop control process, right? Indeed,
the random connection change process would require that some
connections be "unpruned" (connection weights increased from zero) to
move the intrinsic error variable toward zero, right? So the
unsystematic "pruning" done by reorganization can't be like pruning a
tree, where, once something is removed (pruned) it can't really be
re-attached (unpruned). I think that's true, anyway. Correct me if I'm
wrong.

I would still like to know how the pattern recognition process of
pruning works. Maybe it is like E. coli reorganization.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.22.1428 EDT)]

[From Rick Marken (2010.03.22.0920)]

Bruce Gregory (2010.03.21.1752 EDT)--

Rick Marken (2010.03.21.1400)--

BG: That was not my point, but I'm glad you agree. The evidence is that
chess players rely on pattern recognition to prune the possibilities.
Pattern recognition can be described as a closed-loop process, needless to
say.

Could you explain how pattern recognition prunes the (position-move, I
presume) possibilities?

BG: When you recognize a pattern you have seen before in a game, you are likely to recall the successful (or unsuccessful) moves made in that context. This allows you to focus your attention on those moves and ignore other possible moves.

But pruning (eliminating options) implies a control process, one with the
goal of finding the one position-move association (out of the many in
memory) that is "right" (that matches a reference for what the move
should be at that point in the game).

BG: So the player recognizes the correct move because it matches a reference
for the correct move at that point in the game.

No, the reference is the desired (and, thus, "right" from the point of
view of the pruner) position-move. It's what makes the pruning
possible. Without a specification of the position-move desired how
could you know which ones to remove (prune). It's like pruning a tree;
you have to know which branches to keep in order to know which ones to
remove.

BG: I think the question is not how do you make the desired position move, but how to you arrive at the desired position move. If you recall moves that led to desired outcomes, you can focus your attention on similar moves. ("That worked then, will something like it work now?") The controlled perceptions might be something like "recall positions similar to the one I am perceiving now" followed by "explore the outcomes of moves similar to those made from those positions." The actual move made is the one that leads to a pattern characterized as "strong" because similar patterns have led to success in the games studied. In your language, the player keeps the moves that worked in the past and ignores other possible moves. The unpromising moves are not so much pruned as they are ignored; the player is not reminded of them.

Bruce

[From Bill Powers (2010.03.22.1225 MDT)]

Rick Marken (2010.03.22.1120) --

RM: The pruning that is done by reorganization is random (unsystematic)
but it is still part of a closed loop control process, right? Indeed,
the random connection change process would require that some
connections be "unpruned" (connection weights increased from zero) to
move the intrinsic error variable toward zero, right?

BP: "Right" to both.

RM: So the unsystematic "pruning" done by reorganization can't be like pruning a tree, where, once something is removed (pruned) it can't really be
re-attached (unpruned). I think that's true, anyway. Correct me if I'm
wrong.

BP: You're right. But the pruning process that neuroscientists refer to really does take place in infants, and it works two ways: connections can be added as well as destroyed. It's just that at first, after the initial neurogenesis, everything is pretty much connected to everything else nearby, with far more synapses and axons than are found in older children's brains. So reorganization works mostly in the direction of reducing the number of connections. Think of the concepts of neurogenesis and pruning as two localized views of the elephant we call reorganization.

RM: I would still like to know how the pattern recognition process of
pruning works. Maybe it is like E. coli reorganization.

BP: Pruning is one of the first things that happens after the initial proliferation of connections. I don't really see an alternative to reorganization, considering that at first there are no organized higher-order processes. Those are built by reorganization.

Maybe, as PCT becomes more connected with neuroscience, we won't have to do so much guessing. My factual statements about neurogenesis are, as far as I know, guesses based on an inadequate knowledge of even what is known now.
Addictive, but I should try to cut down. Next week for sure.

Best,

Bill P.

[From Rick Marken (2010.03.22.1940)]

Bruce Gregory (2010.03.22.1428 EDT)

Rick Marken (2010.03.22.0920)--

Could you explain how pattern recognition prunes the (position-move, I
presume) possibilities?

BG: When you recognize a pattern you have seen before in a game, you are likely to
recall the successful (or unsuccessful) moves made in that context. This allows you
to focus your attention on those moves and ignore other possible moves.

As I recall that's basically the way Chase and Simon interpreted their
results, which is what led me to see their explanation of chess skill
as S-R. The current chess board position is the stimulus (S) that
leads to a pattern recognition process, resulting in the evocation of
a memory that matches S (current board position). Associated with that
memory - match is a set of possible moves. Attention is focused on
this set of moves which, I presume, results in the selection of the
best move from the set. So the process looks like this:

                                    Attention
                                        >
                                        v
Current --> Memory --> Set of ---> Move
Board Match Moves

    S -------------------------------------------------->R

This is what I call an S-R model because causality moves in one
direction, from board position (S) to move (R), and it is open loop;
the move is the last step in the causal chain; it doesn't loop back to
affect the input. Maybe it would be better to call this an
input-output or information processing model rather than an S-R model
because of all the causal links that presumably exist between S and R.
But this seems to be the model you described above, even though you
have obviously left out a lot of important details (like what
attention is, how it is focused and how this focusing results in an
actual move). Is this right?

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.23.0955 EDT)]

[From Rick Marken (2010.03.22.1940)]

Bruce Gregory (2010.03.22.1428 EDT)

Rick Marken (2010.03.22.0920)--

Could you explain how pattern recognition prunes the (position-move, I
presume) possibilities?

BG: When you recognize a pattern you have seen before in a game, you are likely to
recall the successful (or unsuccessful) moves made in that context. This allows you
to focus your attention on those moves and ignore other possible moves.

As I recall that's basically the way Chase and Simon interpreted their
results, which is what led me to see their explanation of chess skill
as S-R. The current chess board position is the stimulus (S) that
leads to a pattern recognition process, resulting in the evocation of
a memory that matches S (current board position). Associated with that
memory - match is a set of possible moves. Attention is focused on
this set of moves which, I presume, results in the selection of the
best move from the set. So the process looks like this:

                                   Attention
                                       >
                                       v
Current --> Memory --> Set of ---> Move
Board Match Moves

   S -------------------------------------------------->R

This is what I call an S-R model because causality moves in one
direction, from board position (S) to move (R), and it is open loop;
the move is the last step in the causal chain; it doesn't loop back to
affect the input. Maybe it would be better to call this an
input-output or information processing model rather than an S-R model
because of all the causal links that presumably exist between S and R.

BG: I am beginning to appreciate Bill's claim that causality is not a useful scientific concept. Let's assume your description is correct. Your diagram does not stand by itself. Rather it is embedded in a series of loops. The configuration of the chess board does not "cause" my move. If anything causes my move it is the fact that I am engaged in a game of chess. My move in turn is followed by your move which again links back via your diagram to my next move. The fact that I am looking at the current board is part of an activity called "playing chess" which is in turn embedded in a larger goal such as "passing time," "improving my game," or "showing that you are not as smart as you think you are." Pursuing such goals involves purpose and closed loops.

I suspect that your objection is based on the fact that selecting the next move does not appear to be closed loop. However, a test for the controlled variable might show otherwise. If you attempt to prevent me from making the move I have selected, say by grabbing my arm, I will resist the disturbance. Your move may disturb my plan. As a result, I will adopt a new plan because I intend to win or to force a draw.

As I recall, the history of chess playing programs started with efforts to base moves of higher level goals such as "control the center of the board." Those efforts, which may seem to embody closed-loops in a more obvious way, did not prove terribly successful. If PCT can leader to a better approach to programming a computer to play chess, I suspect it will be recognized.

But this seems to be the model you described above, even though you
have obviously left out a lot of important details (like what
attention is, how it is focused and how this focusing results in an
actual move). Is this right?

BG: Since attention is not a trivial topic, I will postpone discussion of it to a later. But for the present, I will say that attention is not a cause; attention is always in the service of achieving goals.

Bruce

[From Rick Marken (2010.03.24.0945)]

Bruce Gregory (2010.03.23.0955 EDT)--

BG: I am beginning to appreciate Bill's claim that causality is not a useful scientific
concept. Let's assume your description is correct. Your diagram does not stand by
itself. Rather it is embedded in a series of loops.

RM: Yes. And this loop is sometimes shown in diagrams of information
processing systems but it's always interpreted sequentially, as a
series of steps. For example, in chess the sequence would be: board
position --> memory --> move selection --> move --> new board position
--> memory etc. Basically a snazzy S-R chain. What they don't get is
the fact that all this is happening at the same time (integrating over
a sufficient time interval) in a negative feedback loop. And when you
have a system (chess player) -- environment (chess board/opponent)
relationship that is organized as a negative feedback loop then the
system (the chess player) is a control system controlling various
perceptual aspects of the environment (board). This is what Simon
seems to have tacitly understood when he developed his chess playing
programs because these programs were organized as control systems. The
chess program made moves in order to achieve goals and subgoals
specified by the program itself.

Simon's chess playing program was built simply to see if one could
build a machine that could play chess; it was a computer science
experiment. The field of cognitive science emerged when it was
realized that these computer programs could be considered a theory of
how people think. Unfortunately, this all happened when the only way
psychologists knew how to test theories of mind was using an
input-output model; the causal model of conventional experimental
psychology. Thus we have the Chase and Simon experiment, which could
only test an input-output model of chess playing: board position
(input) evokes position-move memories (output). So the conclusion of
that experiment was that chess playing is a (sequential) input-output
process: board position --> memory --> move --> new board position -->
etc.

If Simon had understood how control theory applies to behavior -- that
is, if he knew PCT (which he was unlikely to have had much contact
with back then) he would have known that a goal was a reference for an
perception and that the way to test to see whether expert chess
players actually try to achieve the same goals as his chess program is
to test for controlled variables. This would be quite a different
approach to the study of chess playing (and problem solving in
general) than the one dictated by conventional experimental
methodology.

So the essence of the problem (which is why I brought up the Chase and
Simon paper in the first place) is that the study of thinking using
conventional methodology has unfortunate consequences. There are
actually two unfortunate consequences. The first (which I already
mentioned) is that conventional methods virtually force the conclusion
that cognitive behavior is an input-output process since the research
is based on an open loop input-output model. The second is that
conventional methods ignore the possibility that the observed behavior
is organized around control of various perceptual variables:
controlled variables.

Controlled variables are central to the study of living control
systems. The search for controlled variables (using "the test") is
what distinguishes the methods used to study closed-loop control
(i.e., living) systems from conventional methods, which are
appropriate for the study of open loop (i.e., non-living) systems.
Indeed, those with a vested interest in conventional methodology react
to the term "controlled variables" in the same way that a vampire
responds to the presence of a crucifix. The possibility that behavior
is organized around the control of perceptual variables, and that the
input-output relationships that are measured in conventional
experiments are simply the observed efforts by the system to protect
(via output) these controlled variables from disturbance (the apparent
"input"), is the truly revolutionary contribution of PCT to the study
of living systems (in my opinion).

I suspect that your objection is based on the fact that selecting the next move
does not appear to be closed loop.

Not at all. My objection is based on the fact that Chase and Simon
describe chess playing as an open loop input-output process rather
than as a control process. The most obvious evidence of this is that
they don't view chess playing in terms of controlled variables.

However, a test for the controlled variable might show otherwise.

Of course, that's what you would find in fact; but it's not part of
their theory. Of course the placement of the pieces is controlled. But
more important, higher level perceptual aspects of the board are
surely being controlled, such as "control of center","threats to the
queen", etc. These are things that Simon might have tested for if he
had understood control theory.

Actually, there was a study done in about 1980, on solving water jar
problems, that did do something like a test for a controlled variable
when solving these problems. It was a really nifty paper -- by Atwood
and Polson, reported in _Cognitive Psychology_, I believe. What they
found was evidence that one of the variables controlled for when
solving such problems was a perception of the current state of the
problem being similar to the goal state; I think they called it a
"means-ends" strategy. Anyway, what they found is that most
individuals do control for this perception and it's what makes problem
solution difficult. The solution of the problem requires moving the
contents of the jars, at one point in the problem, to a state that
looks _less_ like the goal state. I guess I should have done some
follow up on that research, but other things took priority. But maybe
now is the time, since there seems to be some skepticism about the
relevance of PCT to cognitive behaviors like problem solving.

If you attempt to prevent me from making the move I have
selected, say by grabbing my arm, I will resist the disturbance.

Yes, grabbing you arm is a disturbance to your control of piece placement.

Your move may disturb my plan.

I would say that my move is a disturbance to a perception of the board
that you are controlling for. It is (hopefully, from my point of view)
a disturbance to your control of one perception you are surely
controlling for: the perception of me being checkmated.

As a result, I will adopt a new plan because I intend to win or to force a draw.

Your moves act (like the movements of the mouse in a tracking task) to
compensate for the disturbance my move creates to whatever are the
perceived states of the board you are controlling for. It's those
perceived states that players want to keep the board in that should
have been the main main focus of Simon's study of human chess players.

> As I recall, the history of chess playing programs started
withmhose efforts, which

may seem to embody closed-loops in a more obvious way, did not prove terribly
successful. If PCT can leader to a better approach to programming a computer to
play chess, I suspect it will be recognized.

PCT is aimed at understanding how living systems control and it's
success will be measured by how well it does that. The application of
PCT to chess playing would be aimed at understanding how human chess
players play chess. Since machines can now regularly beat humans at
chess (using exhaustive search algorithms that would be impossible for
a human) it seems unlikely that PCT would be able come up with a
better approach to playing chess than what is already being
implemented in computer programs.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.24.1730 EDT)]

[From Rick Marken (2010.03.24.0945)]

So the essence of the problem (which is why I brought up the Chase and
Simon paper in the first place) is that the study of thinking using
conventional methodology has unfortunate consequences. There are
actually two unfortunate consequences. The first (which I already
mentioned) is that conventional methods virtually force the conclusion
that cognitive behavior is an input-output process since the research
is based on an open loop input-output model. The second is that
conventional methods ignore the possibility that the observed behavior
is organized around control of various perceptual variables:
controlled variables.

BG: Are you saying that the strategy that should be employed in the understanding of behavior is to guess at the perception that is being controlled and then test to see if that conjecture is correct? In that case, we need to guess what perception a chess master is controlling with every move she makes. (of course, if I could do this successfully, I would be a chess master.)

How do you see this description (which I think I understand) as fundamentally different from a description based on goals and ways to achieve these goals? It seems to me that in many cases I can model a goal as a perceptual state. (You and Bill can perceive many things that I can’t, but that is my shortcoming. I can see patterns and patterns of patterns, but nothing much beyond that.)

Controlled variables are central to the study of living control
systems. The search for controlled variables (using “the test”) is
what distinguishes the methods used to study closed-loop control
(i.e., living) systems from conventional methods, which are
appropriate for the study of open loop (i.e., non-living) systems.
Indeed, those with a vested interest in conventional methodology react
to the term “controlled variables” in the same way that a vampire
responds to the presence of a crucifix. The possibility that behavior
is organized around the control of perceptual variables, and that the
input-output relationships that are measured in conventional
experiments are simply the observed efforts by the system to protect
(via output) these controlled variables from disturbance (the apparent
“input”), is the truly revolutionary contribution of PCT to the study
of living systems (in my opinion).

BG: So again, chess masters are acting to preserve certain perceptions from disturbance. we know they do this by making moves. How does this differ from the claim that chess masters have certain goals and they make moves to achieve those goals? (Where, once again, the goal can be specified as a perceptual state.)

RM: Of course, that’s what you would find in fact; but it’s not part of
their theory. Of course the placement of the pieces is controlled. But
more important, higher level perceptual aspects of the board are
surely being controlled, such as “control of center”,“threats to the
queen”, etc. These are things that Simon might have tested for if he
had understood control theory.

BG: One way that expertise has been studied is by asking experts to describe what they are doing. An obvious limitation of this approach is that the experts may be unaware of or mistaken about what they are doing. (I suspect this is true of catching fly balls.) But even when they have an idea such as “protecting the queen” it may be far from obvious how this translates into a controlled perception. It is often easier to identify a goal than to go about achieving it. Student pilots want to avoid colliding with other planes, but need a savvy flight instructor to tell them that they are on a collision course with anything that seems to remain stationary on the windscreen and constantly grows larger. The best way to avoid a collision is often the counter-intuitive step of pointing the airplane at the object. For what it’s worth, knowing how to avoid a mid-air collision does not always mean that you will control this particular perception. Damage to the pathways leading from the prefrontal cortex to the amygdala apparently interferes with the ability to tell whether an event, such as a mid-air collision, is desirable or undesirable.

RM: Actually, there was a study done in about 1980, on solving water jar
problems, that did do something like a test for a controlled variable
when solving these problems. It was a really nifty paper – by Atwood
and Polson, reported in Cognitive Psychology, I believe. What they
found was evidence that one of the variables controlled for when
solving such problems was a perception of the current state of the
problem being similar to the goal state; I think they called it a
“means-ends” strategy. Anyway, what they found is that most
individuals do control for this perception and it’s what makes problem
solution difficult. The solution of the problem requires moving the
contents of the jars, at one point in the problem, to a state that
looks less like the goal state.

BG: This sounds like the “local minimum” problem in an energy landscape. The system gets stuck in a local minimum because it makes energy to get out of a small valley and to continue searching for the deepest valley.

RM: Your moves act (like the movements of the mouse in a tracking task) to
compensate for the disturbance my move creates to whatever are the
perceived states of the board you are controlling for. It’s those
perceived states that players want to keep the board in that should
have been the main main focus of Simon’s study of human chess players.

BG: Can’t we treat every game as a test for the controlled variable? Each player tries to guess the variable the other is trying to control. Every move is an effort to disturb the variable the other player is trying to control. Each player wants to perceive the other as losing. It really doesn’t matter how this comes about, i.e., what the final state of the board looks like, so long as it represents a defeat for your opponent.

Bruce

[From Rick Marken (2010.03.24.2200)]

Bruce Gregory (2010.03.24.1730 EDT)--

Rick Marken (2010.03.24.0945)]

RM: So the essence of the problem (which is why I brought up the Chase and
Simon paper in the first place) is that the study of thinking using
conventional methodology has unfortunate consequences.

BG: Are you saying that the strategy that should be employed in the
understanding of behavior is to guess at the perception that is being
controlled and then test to see if that conjecture is correct?

Exactly! This has been my plea for decades now. I have been trying to
get research psychologists to start doing behavioral research with the
aim of determining whether the system under study is controlling any
perceptual variables and, if so, what these variables are. Both of
these determinations can be made by doing the test for the controlled
variable.

In that case, we need to guess what perception a chess master is controlling
with every move she makes.

Well, I would do the study by first hypothesizing a controlled
variable. This would require defining it, of course. So if you
hypothesize that "control of center" is a controlled variable then you
would have to define what that means in terms of board positions: not
necessarily an easy task. But once you have the hypothetical
controlled variable defined then it should be easy to set up board
positions and know which moves are disturbances to this perception and
which are not. Then you make the moves and see if your subject makes
reply moves that brings her board position back to "control of center"
(as you've defined it); if she's a real good player -- better than you
-- you might not even have anticipated the move she makes that returns
her to "control of the center". But given you definition of "control
of the center" you should be able to see whether or not her move
brings the state of the board back toward "control of center" (for
her).

How do you see this description (which I think I understand) as
fundamentally different from a description based on goals and ways to
achieve these goals?

It's not different at all; you're describing PCT. PCT just makes the
nature of goal achievement more precise. Goal achievement is the
process of bringing a perceptual variable to the reference state
defined autonomously by the system. It is the process of control.

It seems to me that in many cases I can model a goal as
a perceptual state.

I think the word "goal" usually refers to the intended result of
action. So in PCT I would say that the informal term "goal" is better
thought of as the reference signal, which defines the intended state
of the perceptual result of one's actions.

RM: Controlled variables are central to the study of living control
systems...

BG: So again, chess masters are acting to preserve certain perceptions from
disturbance. we know they do this by making moves. How does this differ from
the claim that chess masters have certain goals and they make moves to
achieve those goals? (Where, once again, the goal can be specified as a
perceptual state.)

It doesn't differ at all. Your describing the PCT model of chess
playing, though I find it confusing to think of the perceptual state
as the goal; I think of the perceptual state as the way things are
(like the state "queen threatened") and the reference state as the way
things should be; the goal state (like the state "queen safe").

BG: One way that expertise has been studied is by asking experts to describe
what they are doing. An obvious limitation of this approach is that the
experts may be unaware of or�mistaken�about what they are doing. (I suspect
this is true of catching fly balls.)

Right.

But even when they have an idea such as
"protecting the queen" it may be far from obvious how this translates into a
controlled perception. It is often easier to identify a goal than to go
about achieving it.

You bet! In order to control (bring perceptions reliably to their
reference/goal states) you have to learn to control; that's what
reorganization is about. Chess experts have learned how to achieve
their goals much better than chess novices (like me).I would guess
that much of this learning involves storing in memory rules regarding
what moves to make in order to bring a controlled perception back to
the reference state when it is disturbed in various ways.

BG: Can't we treat every game as a test for the controlled variable? Each
player tries to guess the variable the other is trying to control. Every
move is an effort to disturb the variable the other player is trying to
control. Each player wants to perceive the other as losing. It really
doesn't matter how this comes about, i.e., what the final state of the board
looks like, so long as it represents a defeat for your opponent.

Surely there is some imagining about what state the opponent is trying
to get the board into. So, yes, each time you move you could be
testing to see if the move disturbs what you think the opponent is
controlling. But I think you are also controlling for getting the
board into the position you desire. The move you actually make is
probably one that brings many controlled perceptions closer to their
reference state.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com