Overview
Agents
in MABEL own land configured as parcels. They make decisions on
how to use their land based on inputs from the biophysical environment
(e.g., climate), economics (e.g., anticipated crop prices), demographics
(e.g., age) and any other socioeconomic parameter (e.g., level
of risk adversion). These components are integrated into agent's
beliefs about future expectations about potential payoffs (i.e.,
rewards) from actions. MABEL is a multi-agent based system in that
different kinds of agents, distinguished from their land ownership,
interact in a pseudo market model. Our current market model is
a land bidding routine that allows buyers and sellers of land interact
by first annoucing their intentions and then to complete transactions
given a set of rules, some of which come from a policy agent (i.e.,
a local planning commission), that regulate land use transition
types and how parcels are divided.
Part
of the model involves a complex parcel division routine that allows
buyer and seller agents to calculate three different spatial metrics
related to parcels that would be split. These include a shape and
height-to-width ratio metric. The latter is intended to ensure that
parcels that are created are not "bowling alley" lots.
Bayesian
Belief Networks are used to configure the agent's beliefs linking
biophysical and socioeconomic inputs with internal beliefs and causal
relationships and expected outcomes. We configure the outcomes as
a utility function that allows the agent to calculate its maximum
expected utility given inputs, their beliefs and their intentions.
A learning algorithm is built into the model so that agents can increase
rewards from their actions.
Technical
Dimensions
The MABEL
model is built upon a distributed client-server architecture using
TCP/IP sockets specifically configured to monitor client communication
and simulation status. The server coordinates all of the clients
(so that they are running synchronously). The BBN agent profiles
are kept on the server side as well ensuring that agents across all
client simulations are configured identically.
The model
also integrates code from C, C++, Swarm, scripts from SPSS and ArcGIS.
Swarm is run in the LINUX environment on a Windows simulator. Much
of the data communication software is written in C.
Modeling
Environment
We have
conducted several different types of simulations to test the reliability
of MABEL. This includes a Monte Carlo simulation, tests of different
BBN configurations, tests of different data sources as inputs (e.g.,
PUMs) and tests of different parcelization algorithms.
We are
currently (early 2007) working on developing a set of role playing
simulation games and methods that can be used to parameterize MABEL
for East Africa. This work will entail conducting games that generate
specific outcomes that can be used to configure MABEL in ways that
test our understanding of how people respond to climate change in
savannah ecosystems.
|
