Homework -
Academics choose research topics.
By Fabian
Held and Antonio Miscio
Introduction
We
developed a simple agent-based model of academic research. Our agents are born
as PhD students, graduate into full-blown academic and attempt several research
projects in the course of their academic life. Different factors affect the
choice of what topic to study, some are related to the researcherÕs
characteristics and some with the research environment in which it operates. We
analyse the long run effect of different parameters of interest on the amount
of knowledge discovered, the number of academics still active in research and
the number of initial Òschools of thoughtÓ still surviving after the passage of time.
Theory and
assumptions
When
academics choose which projects to work on they are typicall influenced or
constrained by a number of factors. Our assumptions are intended to represent
what in our view are the most relevant factors characterizing this choice and
can be classified in three areas: knowledge base, academicÕs own
characteristics, peers.
á Knowledge base: new
research topics build on the existing knowledge base
á AcademicÕs own
characteristics: researchers have a field of expertise, i.e. a field on which
they have done some work in the past; they are also born with a natural
inclination for a particular area of research, for instance natural vs. social
sciences; finally academicsÕs choice of topics also depends on factors such as
age, whether they hold a tenured position, how stubborn they are, we condense
all the latter into one factor assuming that they are highly correlated.
á Peers: the research
environment and influence of peers can matter in different ways, for instance
it can act as a signal of what topics ÒsellÓ well at the moment, what is the
predominant paradigm in the discipline, or which recent development are the
most promising.
Moreover,
in addition to working on some projects on their own in some cases academics
team up with one another to produce multidiciplinary research projects, in
which case the resulting topic is a hybrid object.
Finally,
academics are ÒbornÓ as PhD students and at the beginning of their career they
are closely associated with a more senior academic, this has obviously an
impact on the initial topic chosen by the student.
Implementation
We
used Netlogo to model the following elements:
á Knowledge: Knowledge is
represented by a two dimensional grid which is initially unexplored (patches
are initially black), i.e. ideas exist but are not known until an academic successfully
discovers them (patches become white). Each angle in the 0-360 degrees range is
an area of knowledge (e.g. 60 degrees = biology, 95 degrees = anthropology)
á Academics: Academics are
born (more on this to follow) in a location on the grid, somewhere at the
frontier of discovered knowledge. They are endowed with a natural inclination
for a discipline, represented by their ideal angle a; they are aware of the
position of their peers on the grid; they choose research topics and attempt a
project in the direction of the chosen topic; in some cases they generate
offsprings (i.e. little Phds); over time acedemics age and stop doing research;
á Choice of a topic: an
academic starts by observing her own current position; she also observes where
are all the other peers (i.e. agents located within a certain distance from
herself) and calculates in what direction (an angle b) she should walk to reach the average peer position
starting from her current position; finally a topic is selected by taking a
weighted average of a and b, where the relative weight
on the natural inclination increases with age (two interpretations: academics
become more stubborn, academics are less influenced by peer researchers).
á Project: once a topic (i.e.
a direction) has been determined following the procedure detailed above, the
academic attempts a project, i.e. she walks in that direction on the grid until
she finds an area of knowledge which has not been explored yet. Projects are
successful with some probability which depends on the environment, that is the
more knowledge has been explored already in the neighborhood of the new topic,
the more likely it is that a new project on that topic will be successful. When
a project turns out to be successful (cell turns white) the academic can
continue doing research and a new PhD student is generated at that location. If
instead it is unsuccessful then the academic stays on that patch until another
academic successfully explores it.
á PhD students: Even though a
new PhD student is born in the same location as where his supervisor is at that
moment, heÕs endowed his own a, independently of the supervisorÕ. After the first
period, the PhD student ÒgraduatesÓ and becomes a full-blown academic. We also
introduced a parameter to represent the survival rate of PhD students in the
academia because we want to analyse what happens once we allow for a variable
number of them to quit the academia at the end of the PhD and join the private
sector.
á
Multidisciplinarity: a multidiciplinary project is the result of a
random match of two academics, the hybrid projects inherits a trait from each
academic, i.e. the current position of one of them and the preferred direction a of the other one
Netlogo applet starts here
Netlogo applet ends here
Results
We
explored the model's behaviour throughout a sensible range of its parameter
space. The four model parameters PhD retention rate (15%, 20%, 30%, 40%),
initial number of academics (10, 20, 30, 40, 50), probability of
multidisciplinary research (1%, 3%, 5%) and maximum age of academics (5, 7
projects) served as independent variables and the model was run 25 times for
every combination of these parameters. The maximum duration was 70 steps,
because initial exploration showed that at that time the model had either
stopped because all academics had disappeared - or the population of academics
becomes so large that NetLogo slows down substantially.
Dependent
Variables were the maximum amount of knowledge discovered, the number of
academics still actively pursuing research and the number of original ideas or
"schools of thought" that had persisted until the end of the model.
We
used regression analysis to assess the strength of impacts of model parameters
on model development.
All
independent variables turned out to be statistically signifficant, except for
the probability of multidisciplinary research:
Experiment 1: MaxKnowledge ~
PhDSurvival + InitAgents + pMultidisciplinarity + MaxAge
Coefficients: |
Estimate |
Std.
Error |
t
value |
Pr(>|t|) |
(Intercept) |
-11451.79 |
1804.06 |
-6.348 |
4.49e-09
*** |
PhDSurvival |
170.02 |
25.97 |
6.547 |
1.71e-09
*** |
InitAgents |
54.04 |
17.63 |
3.065 |
0.00271
** |
pMultidisciplinarity |
23.43 |
152.69 |
0.153 |
0.87831 |
MaxAge |
1098.66 |
249.34 |
4.406 |
2.37e-05
*** |
Signif.
codes: *** 0.001; ** 0.01; * 0.05.
Residual
standard error: 2731 on 115 degrees of freedom
Multiple
R-squared: 0.384,
Adjusted R-squared: 0.3626
F-statistic:
17.92 on 4 and 115 DF, p-value:
1.831e-11
Experiment 2: Turtles.at.End ~
PhDSurvival + InitAgents + pMultidisciplinarity + MaxAge
Coefficients: |
Estimate |
Std.
Error |
t
value |
Pr(>|t|) |
(Intercept) |
-18430.45 |
3394.79 |
-5.429 |
3.20e-07
*** |
PhDSurvival |
256.80 |
48.87 |
5.255 |
6.89e-07
*** |
InitAgents |
96.17 |
33.18 |
2.899 |
0.004488
** |
pMultidisciplinarity |
65.77 |
287.32 |
0.229 |
0.819352 |
MaxAge |
1721.94 |
469.20 |
3.670 |
0.000369
*** |
Signif.
codes: *** 0.001; ** 0.01; * 0.05.
Residual
standard error: 5140 on 115 degrees of freedom
Multiple
R-squared: 0.3011,
Adjusted R-squared: 0.2768
F-statistic:
12.38 on 4 and 115 DF, p-value:
2.077e-08
Experiment 3: Original.Ideas.Left ~
PhDSurvival + InitAgents + pMultidisciplinarity + MaxAge
Coefficients: |
Estimate |
Std.
Error |
t
value |
Pr(>|t|) |
(Intercept) |
-1.693538 |
0.411095 |
-4.120 |
7.18e-05
*** |
PhDSurvival |
0.022373 |
0.005918 |
3.781 |
0.00025
*** |
InitAgents |
0.010000 |
0.004018 |
2.489 |
0.01424
* |
pMultidisciplinarity |
0.018750 |
0.034794 |
0.539 |
0.59100 |
MaxAge |
0.150000 |
0.056818 |
2.640 |
0.00944
** |
Signif.
codes: *** 0.001; ** 0.01; * 0.05.
Residual
standard error: 0.6224 on 115 degrees of freedom
Multiple
R-squared: 0.1944,
Adjusted R-squared: 0.1664
F-statistic:
6.937 on 4 and 115 DF, p-value:
4.872e-05
Furthermore
the results and dynamics are provided graphically:
Further
research
Given
the limited time available for this homework we had to remove from the model a
number of interesting features which are worth further investigation. In
particular we would have liked to have knowledged represented by a rugged
landscaped instead of a uniform one. The interpretation could be that some
areas of knowledge are harder to discover than others (hence a rugged
environment where cells have a different altitude) and therefore there is a
lower likelihood that a research project is successful in discovering that area
of knowledge.
Although
we modeled to some extent a notion of incremental
knowledge, our patches discretely change from a value zero to one (from
undiscovered to discovered knowledge) as soon a a project is successful. We
could instead imagine that academics, when successful, add a certain amount of
understanding to that patch, and therefore we could have multiple academics
adding understanding to the same patch.
Moreover,
the matching mechanism behind a multidisciplinary project is entirely random, a
more plausible one would instead be conditional on some characteristics of the
academics such as their past record of projects, natural inclination, etc.
Finally,
all projects in this model only last one period and only one project per period
can be attempted, that is an academic attempts a project only once and if she
happens to be unsuccessful then sheÕs stuck there for some time. The model
would gain some more realism by allowing academic to work on the same project
for longer than one period and to work on multiple projects at the same time.
This features can of course be interacted with the project record of the
academic and should depend on the research environment too.