IEEE Technology and Society Magazine - March 2018 - 43

focused on automated hypothesis formation [e.g., 3].
Within the philosophy of science, hypothesis formation
has been closely associated with induction. This focus
may be due to much philosophy of science being physics
focused [16]. In modern biology most hypothesis formation is abductive [11]. What are hypothesized are factual
relationships between entities, e.g., that the gene named
YBR060c encodes for the enzyme with function chorismate mutase, that the gene named YPR060c encodes a
protein with a four-helix bundle topology, etc. Such relationships are factual rather than general laws.
Robot scientists follow a hypothetico-deductive methodology. Hypotheses are formed either using abduction
or induction. The experimental consequences of these
hypotheses are then deductively inferred, and then
physical experiments are conducted to observe what
causally happens in the real world. Adam used abduction to form hypotheses. A set of models is generated,
each with different abduced propositions. With the
model (T) these propositions (H) enable the deduction
of whether growth is predicted (O) for a particular experiment, i.e., T / H : O. These deductions are monitored
by a meta-logical program that determines the truth or
falsehood of the abstract theoretical growth proposition in the various models [19]. This is then integrated
with physical effectors to physically execute an experiment and thereby determine whether actually growth
occurs or not, which can be mapped to the robot scientists abstract model of reality.
Eve uses induction to form hypotheses. To select
compounds to test its hypotheses, Eve uses active
learning [9], [20]. The active learning task is comparable to that in many other areas of science and engineering: identify or design artifacts that have optimal
performance. However, it has an extra ingredient reminiscent of reinforcement learning: balancing the exploration of compound space with the exploitation of
regions of highly active compounds.

Robot and Human Scientists
Is Science Solely a Human Activity?
It is easy to find evidence that science should be viewed
as solely a human endeavor [e.g., 21]. Within the philosophy of science there are many advocates of a humanistic understanding of science. However, developments in
AI question the centrality of human creativity in the creation of scientific knowledge. Although most philosophers of science seem not to have engaged with the
possibility of automating science, the views of certain
philosophers would appear to infer that science, as a set
of practices would be very difficult to automate. Among
these are probably the two best known post-war philosophers of science: Karl Popper and Thomas Kuhn (of
MARCH 2018

∕

course, there are many more authors who have relevant
writings on the topic, e.g., [22]).
We are unaware if Popper ever directly tackled the
question of whether a machine could be engineered to
do science. However, it would seem reasonable to infer
from his other clearly expressed views - that hypothesis formation requires human creativity, and that induction is a myth - that he would have denied the possibility
of mechanizing scientific discovery [10].
In his postscript to The Structure of Scientific Revolutions (1970) Thomas Kuhn responds to critics who
viewed his work as relativistic [23]. Kuhn uses his account
of scientific theories to argue that science is a special
case. This is because viewing proponents of competing
scientific theories as simply akin to members of different
language-communities does not account for scientists as
fundamentally puzzle-solvers [23, p. 205]. Kuhn was an
advocate of scientific progress and believes theories and
paradigms each build upon that which has come
before. Similarly, Kuhn was open to the use of computer programs in scientific knowledge making, and
uses this fact to display his willingness to accept the
importance of rule-following in science [23, p.191]. However, Kuhn's larger point is that it is not quite right to say
that scientists do not follow any fixed rules. He is arguing
that scientists follow rules based on previous exemplars
from their field (see "A Sociological Perspective" section
below). Rules cannot be abstracted from exemplars and
take their place [23, p.192], and it is prior experience
and training that shapes how scientists judge when
rules are being followed or not [23, p. 198]. Applied to
robot scientists the question remains open as to how
prior experience and training can become embedded
in AI programs.
If one accepts that robot scientists can automate
many of the steps in the generation of scientific knowledge, then there would appear to be two main "getouts" that would still enable philosophers to maintain
that science cannot be automated. One get-out is that a
current robot scientists is not aware that it is doing science, and is therefore the robots are not really doing science [24]. This type of argument would also apply to
chess computers [25] - yet we are unaware of any philosopher who has argued that chess computers do not
really play chess.
The other get-out is to deny that what was "discovered" was really novel science. For example, it could be
argued that the new scientific knowledge was implicit in
the formulation of the problem, and is therefore not
novel. The argument that computers cannot originate
anything and can only do what they are programmed to
do is known as "Lady Lovelace's objection" [26]: "The
Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to

IEEE Technology and Society Magazine

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - March 2018