9
The second part of identity is concerned with the initial strategy each Agent
begins with. As stated above, these initial strategies allow us to jumpstart the evolution
of strategies as well as to test assumptions in the simulation universe. These strategies
must compete like any other randomly generated strategy with the only caveat being that
they cannot be replaced. Membership into each strategy group is also genetic, and is
passed from parent to offspring. Because each strategy is allowed to compete freely
against the random and other learned strategies, and therefore does not necessarily reflect
the values of the agent, strict identity grouping are not used as a selection criterion for
violence. This characteristic is more important as a heuristic tool and is very useful in
tracking the evolution and use of strategies in the agentLand simulation.
This highly simplified identity system, while lacking many details or flexibility,
provides a sufficient level of identity to see reflections in the behavior of agents, as seen
in the simulation runs. Later versions of agentLand may reincorporate a more complex
identity system, perhaps coupled with a sexual reproduction system that allows agents to
switch identities and for new identities to emerge. Until that time however, the current
simplified identity system provides enough interesting results to stimulate research and to
guide the next generation of agent identity systems.
The end result of the agentLand AI is a simulation with Agents that can evaluate
their surroundings and choose an action, learn by experience, learn from others, innovate
and differentiate themselves by national identity. These Agents accomplish this through
a combination of an LCFS with social learning and strategy substitution. This process
adds three important elements to the traditional agents used in agent-based international
relations simulations. Clearly the largest benefit is the addition of Agent memory.
Relying on genetics to transfer behavior is inadequate when faced with simulating human
social and political action. Giving an agent memory allows it to remember past
interactions with its environment and with other agents. This added richness adds
considerable sophistication to the level and kinds of interactions that Agents can have
with their environment and with each other. The next important innovation is the ability
for Agents to learn from others. Social Learning adds a whole new dimension to model
making. Because learning occurs between offspring and the parent’s neighbors, dense
population areas become transmission belts of knowledge, transferring ideas between
different Agents.
Finally, the ability to generate new combinations of conditions and verbs adds a
dynamic element to the simulation, introducing the possibility for an agent to develop a
strategy that is superior to the strategies that are assigned to it at the beginning of the
simulation. Agent-Based Modeling combined with the AI of the agentLand simulation
make the difficult task of incorporating ingenuity into a model of resource scarcity
possible. The end result are agents which simulate very important aspects of human
thinking, forming a very useful toolkit for testing behaviors.
Formalizing the Rule Generation Process
Prior research done with using Complex Adaptive Systems to model international
phenomena through aggregation of simple agency rules has held promise. The
fundamental problem with initial approaches to this problem was the ad hoc process of
incorporating rules of behavior into the simulation process. In order to develop more
Convention