Geoscience Reference
In-Depth Information
in their view of the behavioural rules of the agents. This is because minor differences in the rules
can have far-reaching consequences at the system level. In Section 9.3.4, I illustrate sensitivity of the
model's dynamics to variation in agents' behavioural rules with the Schelling model.
Now, the hard part - the rules of agent behaviour in a great majority of the AB models reflect
the personal view of the model developer. Very often, readers easily accept these rules just because
they have nothing to suggest instead. I am not aware of an AB modelling paper (including my own)
in which
all
behavioural rules are based on observations and/or laboratory experiments. Can we
rely on the researcher's imagination, then? A sceptical positivist would, in addition, recall Lee's
seminal requiem (Lee, 1973) and argue that the variety of a system's dynamics is huge, and until
we impose some empirically justified limitations on every agents' behavioural rules, we would just
obtain varying immeasurable aspects of agents' behaviour, along an arbitrarily wide spectrum of
model dynamics.
The aforementioned discussion is often cited as a conflict between exploratory and predictive
modelling (Batty et al., 2012), where a wider and deeper view of it can be found in Edmonds and
Moss (2005) and in several papers of the recently edited volume on AB modelling in geographical
systems (Batty, 2012; Batty et al., 2012; O'Sullivan et al., 2012). My view is that we have to delay this
discussion until the theories and experimental results of behavioural science are fully incorporated
into geographic AB modelling. The paper of Kennedy (2012) in the aforementioned volume, as well
as the recently edited volume on heuristics (Gigerenzer et al., 2011), can be starting points for those
students of AB modelling that are interested in a tight connection between AB modelling and behav-
ioural science. For now, our knowledge of the quantitative aspects of human behaviour is limited,
and a model as a tool for studying the consequences of possible human and institution behaviours on
the socio-spatial dynamics of the systems is, often, the only way to quantify these dynamics.
9.2.4 t
eMPoral
r
eSolution
of
a
gent
B
ehaViour
Intuitively, the developers of every GC model start with the
discrete time
view: time in the model
is advancing by some constant interval - year, month, day or hour (Benenson and Torrens, 2004,
Chapter 5), and at each time step, some of the agents make decisions that influence other agents
and objects at the same or future time steps. Conceptually, this intuitive view does not fit the AB
approach; the latter, by definition, aims at registering every decision of every model agent, as and
when that decision is made. Computer science resolves this contradictory problem by the
event-
driven
approach to model development (Faison, 2006). The essence of this view is in managing
time as a continuous variable. The agent performs an action depending on the time that has passed
from the moment of the previous or anticipated agent's action or, more generally, any system events.
For example, a driver cancels parking search and drives to a paid lot 2 min before the start of the
business meeting.
In case the modeller can decide on the minimal time interval between system events, the event-
driven approach can be easily implemented with the standard discrete time scheme. Usually, the
minimal model time interval, as dictated by the rules of agent behaviour, is short relative to the dura-
tion of the period of time during which the modeller wants to observe the dynamics of the system.
AB models where agents relocate in space over short time intervals demand a high spatial resolu-
tion. Let us consider the PARKAGENT model as an example. A driver agent in the PARKAGENT
model must recognise whether a parking place is free or occupied. A time interval during which
the driver passes the parking place may thus be a natural time step for the PARKAGENT model.
Field research provides an estimate of the typical speed of the driver that searches for parking -
10-15 km/h, that is, 3-5 m/s (Benenson et al., 2008). The length of the parking place is about 5 m,
and, therefore, a time step of 1 s is sufficient for an unambiguous description of the parking agent
behaviour. For the aforementioned example of horticulture in the African savannah, the natural
time step is a year: a farmer decides on the use of the field and on the future crop once a year regard-
less of whether the decision is made a week before or later.
