CSCE 475/875
Final
Project Assignment:
Local
Decisions vs. Global Coherence
Assigned: September 15, 2011
Due: December 14, 2011
(Project 5
minutes late will not be accepted)
Introduction
The goal of this final
project is to learn how to design a multiagent system (MAS) where agents make
local decisions leading to coherent emergent behaviors that satisfy the design
objectives. Through this exercise, you
will be able to learn to appreciate the complexity of such a design as well as
the simplicity of the resultant design once it is developed. The key is that we know that in general more
global directives and more global-centric decisions will lead to better
coherence more easily than very local-centric decisions and very little global
control. However, a MAS with more global
directives and more global-centric decisions reduce the autonomy and
flexibility the individual agents. When
agent autonomy is reduced, the unique benefits of having an agent-based
solution are also reduced. This has been
one of the most important problems that MAS researchers have tried to work ever
since the inception of the MAS research area.
In this final project
assignment, you will submit a proposal for your final project, use the Repast
simulation software, run comprehensive experiments, and report on your design
and results.
* If a team intends to conduct a study different from the “local
decisions vs. global coherence” objective, I am open to that but I do require
the team to argue and justify their choice, to make a compelling case.
Repast Simulation Software
Check out http://repast.sourceforge.net/index.html for a detailed description of the testbed. There is also FAQ at the website – please review them carefully. In particular, please look at Repast Simphony. A detailed demonstration of Repast environment will be given in the classroom on September 20, 2011.
Briefly, as reported on the website: The Recursive Porous Agent Simulation Toolkit (Repast) is one of several agent modeling toolkits that are available. Repast borrows many concepts from the Swarm agent-based modeling toolkit. At its heart, Repast toolkit version 3 can be thought of as a specification for agent-based modeling services or functions.
Repast has a variety of features including the following:
· Repast includes a variety of agent templates and examples. However, the toolkit gives users complete flexibility as to how they specify the properties and behaviors of agents.
· Repast is fully object-oriented.
· Repast includes a fully concurrent discrete event scheduler. This scheduler supports both sequential and parallel discrete event operations.
· Repast offers built-in simulation results logging and graphing tools.
·
Repast has automated
· Repast provides a range of two-dimensional agent environments and visualizations.
· Repast allows users to dynamically access and modify agent properties, agent behavioral equations, and model properties at run time.
· Repast includes libraries for genetic algorithms, neural networks, random number generation, and specialized mathematics.
· Repast includes built-in systems dynamics modeling.
· Repast has social network modeling support tools.
Design and Implementation
You must design an agent that makes local decisions through communicating (directly and/or indirectly) with other agents in the multiagent systems. The goal is to minimize global control, keeping those directives to a minimum, and yet still maintain a coherent, emergent behavior that satisfies the design objectives of the system. Thus, you must be able to articulate the motivations for an agent to make the correct local decisions and how. Then, you must be able to show that if every agent makes the correct local decisions (and they will since they are all motivated to do so), then the desired overall outcomes can be achieved. Since a MAS lives in an environment, you also need to specify the environmental parameters that will in turn help you design your agent.
This is not easy. Thus, you will probably have multiple agent designs, conduct numerous experiments, and iteratively refine your agent design and your environmental factors.
You are required to turn in only one demo-ready agent in Java for every type of agent in your environment. For example, if you are simulating an auction scenario, your environment would include an auctioneer and a set of buyers. In that case, you need to turn in one auctioneer agent and one buyer agent. Your simulation program then would replicate that buyer agent to generate the required set of buyer agents.
You should also make sure that your Java programs run on Windows. For example, if you write your program in Linux, and port it to Windows, you may see some extra weird characters attached to the end of each line of the code. Such characters may prevent the windows environment from running your program correctly.
Experiments and Report
You are required to run an experiment with your agent design. You must propose a set of hypotheses that you want to validate. And then you must design a study that will allow you to collect the data that you need to validate or invalidate the hypotheses. It is likely that you will have many different environmental settings and different agent designs. Systematically evaluate them and report on the results. For your report, you must provide POJI of the results: Presentations, Observations, Justifications, and Implications. We will discuss this further in class.
It is important that your experiments discuss how you achieve coherent, emergent behavior via local decisions. In particular, show the empirical data on communication, coordination, and so on, that support your claim that you minimize global control in local decision making and still achieve coherent, emergent behavior. Of course, this is a “range”. And thus, degrees of global control and the quality of the emergent behavior will have to taken into account in your discussion.
Notice that the description of simulation experiments should summarize all of your trial simulation runs. This summary should include: (a) The Simulation Parameters (b) The Initial and Final Values of the Environment Variables (that you want to observe), (c) the Random Number Generator and the Seed (do not use timestamps, use constant long integers for each run), (d) Description of the Participating Agents (what type and how many), (e) Summary of the Initial and Final State of Each Agent (e.g., the amount of money an agent has in an auction simulation) (e) Total Tick Counts, and (f) The Final Emergent Behavior. In brief, your simulation summary should allow us to replicate your trial experiments by running your submitted code. Note that your proposal will essentially be a part of your report. So, the better you do your proposal, the less you will have to do towards the end of the semester when you write up your final report.
Final Project Demo Day
On the Final Project Demo day, each team will present their agent and work. Each team is also required to introduce (~5 minutes) its team before the demo, including the overall strategies. This is to let all students know about the different approaches chosen.
Team
You are required to form a team of two or three students and give a name
for your team. All members of the team
get the same score. (NOTE: This team WILL be the same team that
also participates in the Game Days!)
Important Dates
October 15: A sufficiently detailed proposal that describes(1) the problem statement, (2) the agent design strategy, (3) the desired emergent behavior, (4) the hypotheses, and(5) the experiments that you will likely conduct.
Dec 1: Last day to submit your final agent design so that we can start running them to collect statistics on local decisions vs. global coherence. (Note: This gives you at least 14 days to do your experiments and POJI.)
Dec 14: Demo Day!
Grading
The final project will be graded in 2 parts: programming (50%) and report (50%). The programming part will be graded based on: (a) 45% Program Correctness, (b) 15% Software Design, (c) 10% Programming Style, (d) 15% Testing, and (e) 15% Documentation.
The report will be graded based on: (a) 50% Design Description and Discussion, (b) 20% Organization, (c) 10% Requirements, (d) 10% Description of Simulation Experiments, and (d) 10% Grammar and Errors. The report must be written in a “manuscript format” compliant to AAAI, ACM, or IEEE. These manuscript formats can be found online at:
· AAAI: http://www.aaai.org/Publications/Author/author.php
· ACM: http://www.acm.org/pubs/submissions/submission.htm
· IEEE: http://www.acm.org/pubs/submissions/submission.htm
The quality of the design will be
based on our own testing of your software.
The TA will collect the following data on your local decisions: (a) Number
of Agents, (b) Initial Values of the Environment Variables, (c) Initial and
And we will collect the following data on the convergence rate of your system and how close the outcome performance matches the expected outcome: (a) Simulation Parameters, (b) Number of Agents, (c) The Final values of the Environment Variables, (d) Emergent Behavior of the System, and (e) Total Tick Counts (i.e., number of timesteps).