Installed: 8 Jul 2014
Last updated: 8 Jul 2014; 31 Oct 2014
This document is
Scientists aim to discover causes either for the advance of knowledge or for the
sake of improved prevention and control, e.g. of diseases, crop failures,
climate change, and other processes, or production of new phenomena, e.g. safe
nuclear power sources.
Engineers, including civil, mechanical, electrical and computer engineers, make
use of their understanding of causal mechanisms when designing new materials,
machines or structures, extending old designs, or identifying and eliminating
design flaws when things don't work as expected.
Psychologists and biologists investigate the understanding of causation and uses
of causal reasoning in humans, including young children, also in other animals,
while neuroscientists attempt to understand the causal powers of brain mechanisms
that support those and other forms of reasoning and learning.
Roboticists concerned with producing intelligent machines need to understand how
to give their robots the ability to discover and use causal relationships.
Computer scientists attempt to design new forms of computation and new
formalisms making it possible for programmers and computer systems designers
to produce machines with new sorts of causal powers, e.g. resistance to failures
caused by hardware faults, and other forms of robustness.
Can all these and researchers in other disciplines share their knowledge, and
their puzzles, about causation, for mutual benefit?
Many non-philosophers have wondered about, or worried about, whether they can
ever make decisions freely if everything that happens is caused, and, if not everything
is caused, whether uncaused processes must be random and therefore not exercises
David Hume (among others) challenged the claim that we have any clear notion of
causation other than the notion of a learnt regularity that we use to predict
what will come after some observed event or state of affairs. He suggested that
the notion that there is something more than observed correlation, e.g. some
necessary connection, is pure myth, and many contemporary philosophers think that
the notion of an instance of a reliable correlation is the only concept of
causation that we have.
Immanuel Kant, disagreeing with Hume, argued that the requirement that all
concepts must ultimately be based on experience of instances would not allow any
learning to get off the ground, since experience requires use of concepts. (He
therefore disposed of what is now called "symbol grounding theory" in 1781.) He
also claimed that in order to think about a kind of reality that exists
independently of our perceiving it, we need to assume a kind of causal necessity
that is stronger than mere experienced regularity, as a feature of that reality.
Kant's claim can be illustrated using examples of events that have mathematical
consequences. These are situations in which a change in certain properties or
relationships necessitates changes in other properties and relationships.
The Konigsberg Bridge Problem
The configuration of seven bridges linking two islands in the river and the
surrounding town causes the impossibility of traversing all the bridges in a
single tour, without crossing any bridge twice. Convince yourself that it is
impossible, and that either adding or removing a bridge will cause a tour
traversing every bridge to become possible. Does it matter what material
the bridges are made of? See the Wikipedia entry.
Can a baby learn to reason mathematically about the effects of rotating
one rigid impenetrable gear wheel meshed with another?
That Kantian "necessary connection" view contrasts with a modern variant of the
Humean "mere observed regularity" view. This modern variant uses a notion of
"possible world", and relationships between possible worlds. So our complete
universe as it is, past present and future, is one possible world, but there are
other possible worlds in which different things happen or exist. In this framework
(Possible worlds semantics) talk of causal connections is interpreted as
referring not only to observed regularities but to regularities that would have
been preserved even if different initial states or surrounding circumstances had
existed. So on this neo-Humean analysis the truth of A caused B depends on
whether B would or would not have occurred in other possible worlds in which A
does occur and other possible worlds in which A does not occur.
In contrast, a "power" theory of causation claims that the possible worlds story
(which may or may not be coherent, depending on what sort of things possible
worlds are supposed to be) does not explain what causation is. It merely
summarises some of the consequences of causal relationships. These
debates have implications both for theories about how causal understanding works
in humans, and other animals and how it will need to work in future intelligent
machines. But also implications for what the universe needs to be like for
causal connections to be able to exist. There are different variants of the
One of the reasons for inviting computer scientists and roboticists to this
meeting is that there are interesting problems about causation in virtual
machines, and problems about how intelligent machines need to understand
causation, both in themselves, and in their environments.
For more on theories about causation see this background document:
For presentations on different kinds of causal understanding (in particular
Kantian and Humean causal thinking) in humans and other animals see the linked
presentations by Jackie Chappell and Aaron Sloman at the WONAC Workshop in
Oxford in 2007
WONAC: International Workshop on Natural and Artificial Cognition
Pembroke College, Oxford, June 25 and 26, 2007
[TO BE EXPANDED]
Unfortunately most philosophical discussions of these topics seem to ignore what
computer systems engineers have discovered and developed in the last half
century or so about Virtual Machine Functionalism, discussed in
In particular most philosophical discussions seem to ignore what has been
achieved by engineers in giving machines, especially machines that include
sophisticated virtual machinery the ability to have, and to use
intentional states, for example when attempting to identify faults in complex
machinery, or when controlling the behaviour of a complex machine (e.g. landing
an airliner, or monitoring flows in oil refineries. (In the UK there was a major
research project on transferring responsibility for such tasks from human
operators to AI systems, in the 1980s, as part of the Alvey Project.
Direction of fit
Philosophers (e.g. Elizabeth Anscombe) have distinguished the "direction of fit"
of intentional states to what they refer to in belief-like states, and desire-like
states. The same distinction between different "directions of fit" is important
in many control systems -- that was John McCarthy's point in saying that a
thermostat can be "fooled" into turning off a boiler by holding a candle under
Instead of arguing in a knowledge vacuum about what machines (especially
computers) can and cannot do philosophers need to learn about the many
intermediate cases and what difference they make, e.g. cases between a
homeostatic control system and a plant control system or a future intelligent
tutoring system, which will need to be able to refer to mental sates and
processes in learners.
Present some of the "standard" Philosophical background, extended with examples of "computational causation" (e.g. causation in virtual machines composed of multiple asynchronously interacting virtual machines) and "mathematical causation" (e.g. moving a vertex of a planar triangle further from the opposite side causes the area to be increased, adding three marbles to a box containing five marbles causes the number of marbles to go up to eight, changing the curvature of a line causes infinitely many distances between parts of the line to change.____________________________________________________________________________
Example design issue: Reward-based vs Architecture-based motivation
School of Computer Science
The University of Birmingham