Systems science has been instrumental in breaking new scientific ground in diverse fields such as meteorology, engineering and decision analysis. However, it is just beginning to impact public health. This seminar is designed to introduce students to basic tools of theory building and data analysis in systems science and to apply those tools to better understand the obesity epidemic in human populations. There will also be a lab in which students will use a simple demonstration model of food acquisition behavior using agent-based modeling on standard (free) software (netlogo). The central organizing idea of the course is to examine the obesity epidemic at a population level as an emergent properties of complex, nested systems, with attention to feedback processes, multilevel interactions, and the phenomenon of emergence. While the emphasis will be on obesity, the goal will be to explore ways in which the systems approach can be applied to other non-communicable diseases both nationally and internationally. Topics will include: a) the epidemiology of obesity across time and place, b) theories to explain population obesity, c) the role of environments and economic resources in obesity c) basic concepts and tools of systems science, d) modeling energy-balance related behaviors in context, e) agent-based models, systems dynamic models, and social network models
Development of a Simple Model with Environment, Peer Networks, and Agent Behaviors
Loading...
来自 Johns Hopkins University 的课程
Systems Science and Obesity
97 个评分
从本节课中
Module 4
Agent based modeling and conceptual motivation for agent based modeling. This module includes a lab lesson.
与讲师见面
Jessica Jones-SmithAssistant Professor
Human Nutrition
Rahmatollah BeheshtiPostdoctoral Fellow
Department of Civil Engineering
Daniel TaberAssistant Professor
School of Public Health (UT Health)
Kayla de la HayeAssistant Professor of Preventative Medicine
Keck School of Medicine (USC)
Thomas GlassProfessor
Epidemiology
Hello, this is Tak Igusa and
I'm a Professor of Civil Engineering at the Johns Hopkins University.
In this lecture, I want to show some basic concepts about Agent-based Modeling.
So here, I begin with the most basic components of an agent-based model.
First, we have an agent shown on the right.
An agent is a computer simulation of a person.
And on the left, I show the Environment, which in this case,
would be a part of a neighborhood.
And what we would do is we would embed this Agent into the Environment.
So in this slide, we show two types of people.
On the top are people with Unhealthy habits and
on the bottom there are people with Healthy habits.
And these people will be embedded into two types of neighborhoods.
On the top is the Obesogenic neighborhood and
on the bottom I show a Non-obesogenic neighborhood.
So, in an Obesogenic neighborhood,
is one where it is difficult to find healthy foods.
And in a Non-obesogenic neighborhood,
these healthy foods are more accessible to the people living in that neighborhood.
Here, I show the effect of the environment on the agents.
So we begin with the agents on the left side of the slide,
where we have one agent in the obesogenic neighborhood on the top.
And an agent with a non-obesogenic neighborhood on the bottom.
So, over the course of time,
the environment has influences on the agents' behavior.
So on the right side of this slide, you can see how the agent who was initially
healthy in the obesogenic environment has changed to unhealthy behavior.
Were is the agents in the bottom half of the slide,
living in a non-obesogenic neighborhood would be able to retain healthy habits.
Here we see, the agents in a elementary agent-based model.
This was programmed in NetLogo.
In NetLogo, we have some limitations on representing the agent,
so the agents here are shown in blue and red.
And this is as much as we can do with the NetLogo,
to represent the two types of agents, the Unhealthy habits and Healthy habits.
We understand that this may be difficult for some people to see but
this is unfortunately, a limitation of the model.
But in these slides, I will explain everything in the model, using words.
So in the model, there are two types of neighborhoods, on the left portion
of the slide is a square representing the neighborhoods in the agent-based model.
In the upper right portion of that square, is obesogenic neighborhood and
in the remainder of the square of the non-obesogenic neighborhood.
We will embed the agents into the model and
at the beginning the agents are scattered with various types of habits.
And then we will see how the environments will effect the eating habits
of the agents.
If we let the agent-based model run,
then the effect of environment on eating behavior is quite simple.
In the Obesogenic neighborhoods, on the upper right portion
of the model., all of the agents adopt Unhealthy habits.
Whereas in the rest of the neighborhoods, the Non-obesogenic neighborhoods,
all of the agents adopt Healthy habits.
In this slide, we include randomness in the agent behavior.
By including randomness in the agent behaviors,
what happens in the model is that about 10% of the agents
living in the Obesogenic neighborhood are able to maintain Healthy habits.
Whereas, in the Non-obesogenic neighborhoods,
there's about 10% who have Unhealthy habits.
If we look at the results of our model.
One thing we might find is that we could get very similar results
using stand statistical models.
So the question here might be, why are we bothering with agent-based modeling?
And I hope with the next slides, I can show what are the advantages and
what are the additional insights that can be gained with agent-based modeling.
One of the main additional features of agent-based modeling is that it allows
the modeller to endow individual behaviour at the agent level.
So here, I want to show an example of this type of individual behavior.
We begin with a system of rewards, so on the top,
we see a figure which is our agent who has unhealthy habits.
And we will assign no points for that behavior.
On the bottom, there's an agent who adopts healthy behavior.
And that agent would be assigned a number of points,
which we denote as R sub healthy and for clarity, we'll just use 5 points.
Now I should mention that these points are not the points assigned
by someone as an observer looking at an agent.
But these are the points that each agent assigns to themselves.
So in other words, an agent who adopts healthy behavior,
gets 5 points just because of the health benefits of those behaviors.
Next, we show the effect of the environment on our system of rewards.
On the top side of the figure, we see an agent with healthy behavior living in
a non-obesogenic and direct environment and that agent will still have 5 points.
On the bottom half of the figure,
we see an agent with healthy behavior living in obesogenic environment.
We want to put in, some sort of effect of this environment
on a person trying to adopt healthy eating behaviors.
This effect should be negative because if it's difficult to
find healthy foods in this environment, it may be more costly.
The agent may have to travel further distances.
Maintaining a healthy lifestyle in this obesogenic environment
is more difficult because there are maybe more distances to travel
to obtain healthy foods and maybe more costly to obtain these healthy foods.
So we show this negative effects as an environmental effect of negative 3 points.
So, that means an agent trying to adopt a healthy behavior and
then obesogenic environment has a net of only two points.
Next, we show the rewards for following the behavior of the peers of each agent.
So in the top side of the figure, we see agent with
unhealthy behaviors, and that agent has four friends, or peers.
And each of those friends adopts healthy eating behaviors.
So that means the agent at the top has no peer that
the same behavior, so there are no points in this case.
On the bottom side of the figure,
we replace the agent in the middle with an agent with healthy behavior.
And the healthy behavior is in a sense reinforced by these friends.
Each of these four friends also have healthy behaviors.
To show this reinforcing behavior, we assigned two points for
each tier, that shares the same behavior.
So on the bottom figure, the agent gets an additional 8 points from the tiers.
In this slide, we use the same agents.
In the middle, but we placed the peers by four agents all with unhealthy behaviors.
So now the points are switched.
The agent at the top in the middle, has unhealthy behaviors and
all four of the agent's friends, also have unhealthy behaviors.
So that agent has 4 times 2 or 8 points.
The agent on the bottom, has healthy behaviors in the middle but
all of that agents peers, has unhealthy behaviors.
So since none of the peers share the same behaviors as the agent in the middle,
there are no points assigned.
In this slide,
we show what happens when the peers do not all have the same behavior.
As in the previous two slides, the agent in the middle on the top half of the slide
has unhealthy behavior, and the agent on the bottom has healthy behavior.
But unlike the previous two slides, there is a mixture of agents in the peers.
There are 3 agents who are peers that have unhealthy behaviors and
1 agent with the peer that has healthy behaviors.
So on the top half of the figure, we see that the agent in
the middle shares the same unhealthy behaviors as 3 peers.
So the net reward is 3 times 2 or 6 points.
On the bottom half of the figure, the agent in the middle has healthy behavior,
and only 1 peer shares the same healthy behavior.
So the number of points 1 times 2, or 2 points.
Here, we summarize the rewards for
the various eating behaviors and environments.
So we start at the top line, the baseline is where we have
healthy eating behavior, and the reward is 5 points.
On the second line, we see the effect of the obesogenic environment,
that has a negative effect of minus 3 points.
Further down in this table, we see the interactions from the peers.
So a healthy agent interacting with healthy peers gets 2 points per peer.
Whereas an unhealthy agent,
interacting with unhealthy peers, also gets 2 points per peer.
Now, here we introduce the effect of an intervention.
So at this point, we have not yet included the influence that
people in public heath might have on the eating behaviors of the agents.
And we allow this influence in our model by including the effect of an interaction
and then interaction would have an end effect of 2 points for
those agents who adopt healthy eating habits.
In the next few slides, I will show you examples
of all of this point structure for rewards.
In each of these slides, we show an agent in the middle adopting either
an unhealthy habits on the top or healthy habits on the bottom.
We also show four peers, three peers who are adopting
unhealthy habits and and one peer adopting healthy habit.
And we combine all of the points shown on the previous table for each scenario.
So on this slide, the agent should adopt healthy behavior, as shown by the arrow.
Why is that true?
This is because of the point structure.
So we begin by adding the points for the agent who adopts unhealthy behavior.
There are zero points for unhealthy behavior but
there are three peers who also have unhealthy behavior.
So there are 2 times 3 or 6 points for the influence of the peers.
There are zero points for the environmental effect.
So, the net number of points for the agent at the top, is 6.
On the bottom, we consider the same agent but
with healthy, rather than unhealthy behaviors.
There's several things that change in the point structure.
First, there are five points for the healthy behavior.
But there's only one peer that also shares the healthy behavior.
So there's only two times one or two points for the peer effect.
And in this case there is zero environmental effect.
So the net number of points is 7 and so 7 on the bottom half of the figure is
greater than 6 on the top half the agent adopts healthy eating behaviors.
Here, we show the effects of the obesogenic environment
on the behavior of the agent in the middle.
We should remember that the agent in the middle will always choose the behavior
that gives that agent the most number of points.
So in this slide, what happens is that the environment is now obesogenic and
that has a negative effect on a agent who tries to adopt healthy behaviors.
So if we compare this slide from the previous slide,
the only difference is on the agent on the bottom.
Where there is a negative 3 points for the environmental effect,
so you may recall, that an agent trying to adopt healthy behavior
in an obesogenic environment would have a penalty in effect of 3 points.
So [INAUDIBLE] here, where the agent in the bottom, who is trying to adopt healthy
habits has three points subtracted from his reward structure.
So that, while that agent initially had seven points on the previous slide,
the agent now has only 4 points.
So if you compare the number of points for adopting either healthy behavior or
unhealthy behavior then the unhealthy behavior wins with 6 points.
So that means, the agent will adopt unhealthy behavior.
Here, we see what happens if we introduce an intervention.
And we model this intervention as a change in the environment.
So as shown in the table with the reward points structure,
the intervention assigns two extra points to any agent who adopts healthy behavior.
So what happens here is that, while that agent on the top adopting unhealthy
habits still maintains 6 points for that behavior.
All because of the share behaviours from the three piers.
The agent on the bottom, has five points for the healthy behaviors,
two points because of one peer that shares that behavior.
And an additional two points from the intervention, so
the net number of points is 9, which explains why the agent
in this intervention environment will adopt healthy eating habits.
So, here we return back to the agent-based model.
We initialize the model, so that about half of the agents
adopt unhealthy habits and half adopt healthy habits.
And we place them in the Obesogenic and Non-obesogenic environments.
And then we see what happens to the agent behaviors because of these environments.
So here, we see the effect of the environment on the agents.
In the obesogenic neighborhood,
there’s only 10% of the agents who maintain healthy habits,
whereas the remaining 90% now have unhealthy habits.
Where is in the Non-obesogenic neighborhoods,
about 80% are healthy and 20% are unhealthy.
At this point, we include our intervention which is in the very
top right corner of the neighborhoods represented by the agent-based model.
So here, we see what happens when we run the agent-based model
with the intervention zone on the upper right corner.
First, we notice that almost all of the agents
in the intervention zone have adopted healthy eating habits.
But the changes in eating habits propagate away from that intervention zone.
And that is shown by the dark squares
that are dotted throughout the upper portion of the model.
This agent-based model shows the influence of the pure effect.
What happens is that the agents in the intervention zone
are able to influence agents outside of that zone
because they have peers that live outside of the zone.
And because there's a point structure associated with sharing the same behavior.
And so what happens is, because of this concentration of
agents with healthy eating behaviors in the intervention zone.
They're able to influence their peers outside of that zone through
their peer influences and the associated reward structure.
That influence propagates sequentially first to
the agents immediately outside of the intervention zone.
And then subsequently through the interactions from these
peers to agents even further away from the zone.
The net effect is shown in those black squares in the agent-based model.
In each of those black squares, there's an agent who has changed from
unhealthy behavior to healthy behaviors because of the peer effects.
And we can see that those effects have propagated
quite far from the initial intervention zone in the upper right corner.
I hope these slides a good foundation for
the basic understanding of agent-based modelling and its capabilities.
