Hyper-Economy Development Group: software model and project plan
Why banner exchange as the first implementation
While the model itself remains generic, we have chosen banner exchange as
the first prototype implementation for our developing hypereconomic
system for several reasons:
- I have experience with similar systems (ACF)
- this is a good example of already existing multi-agent system
with data practically available, and with very complex
considerations (user preferences) producing numerous points
of simple formal data (page views and banner clicks)
- the system can be expected to be practically useful in this
domain as soon as it's ready.
- TrafficX developer Joel Shellman is willing to implement it
I also hope to further develop the system and see it implemented
in other domains.
Project plan
The first stages of the project will be implemented in Java on Win'95
For development and testing, a model of the domain will be used.
Why a model? - because it allows multiple runs of the system
without bothering real users, until things are ready.
Parts of the system
(more detailed descriptions to follow soon)
- domain model:
data representing agents of different kinds and their
environment. In the first implementation, that will
be pages, banners, their categories and other parameters.
- Event model
Simulates "natural events" in the system, i.e. things that
happen regardless of the agents' actions. In a banner
exchange system, these events are represented by page views
by Web users. The event model will generate these events
with frequencies and fluctuations similar to those in the
real system.
- System decision model
This is the brain of the system, representing collective actions of
agents it will read the state of the environment and generate certain
actions. In a banner exchange, it will decide what banners to show
where and when. The current implementation has five models:
- "Random" - a banner is selected randomly out of banners with maximal inventory
- "Schedule" (utility) - every site agent chooses the [generally] best banner
out of currently available.
- "Economy", with sites and banners bargaining with each other for the opportunity
of banner views, based on their individual experiences.
- "Intermediate" - The economic decisions are made with some general knowledge on
resource utility
- "hyper-economy", with sites and banners sharing their situational
experience with each other, to find the best matches without testing
every banner on every page.
At this point the models work in a simple environment with stochastic events;
the distribution patterns of events do not change in time. The next stage
should involve fluctuating environment.
- Environment response model
The displayed banners will be "clicked" on or not, depending on a set
of realistic parameters (not known to the decision- making module)
- assessment of decision quality
a set of routines showing the system's performance statistics,
such as average click rate [per category], speeds of adaptation
under different conditions, etc. This will allow us to see what
works better - a centralized system, an economy, knowledge
exchange, or their combinations.
After the models are tested, the real-world implementation will
use TrafficX instead of the event model and system response model.
The other parts will stay.
Last updated January 10, 1999.
Back to hypereconomy page