added a new example: city walking behaviour

[Sahil-Chhoker]

Add Walking Behavior Agent-Based Model

Summary

This PR introduces an Agent-Based Model (ABM) for simulating walking behavior in a hypothetical city. The model explores how socioeconomic status (SES), built environment, and social factors dynamically influence walking patterns. The following files and components have been added to the repository:

---

Files Added

1. README.md:

   • Provides a detailed overview of the model architecture, agent characteristics, environmental layers, and simulation scenarios.

2. Model.py:

   • Sets up the grid and initializes the four scenarios (RR, RS, CR, CS).
   • Manages the safety and aesthetic property layers.
   • Places different workplaces according to the scenario.
   • Places Humans according to their SES values
   • Collects metrics such as average daily walking trips, work-related trips, and leisure walks for analysis.

3. Agents.py:

   • Demographics:
     • Gender (50/50 distribution)
     • Age (18-87 years, random distribution)
     • Family Size (1-2 members)
     • Pet Ownership (20% dog ownership rate)
   • Personal Attributes:
     • Walking Ability
     • Walking Attitude
     • Employment Status
     • Social Networks
   • Workplaces:
     • Grocery Store
     • Social Place
     • Non-Food Store
     • Others

---

Key Features Added

1. Initialization Parameters

• Grid Dimensions: Width and height of the city grid.
• Workplace Distribution: Configurations for Grocery Stores, Social Places, and other workplaces.
• Population Demographics: SES-based household positioning, age, family size, and pet ownership distribution.

2. Environmental Layers

• Safety Layer (safety_cell_layer): Dynamic safety values according to the scenarios influencing route selection.
• Aesthetic Layer (aesthetic_cell_layer): Center-focused aesthetic value distribution impacting walking preferences.

3. Simulation Scenarios

• RR (Random-Random): Random distribution of land use and safety values.
• RS (Random-Safety): Random land use with lower safety in core areas.
• CR (Centralized-Random): Centralized land use with random safety values.
• CS (Centralized-Safety): Centralized land use with lower safety in core areas.

4. Agent Characteristics

• Agents include attributes like walking ability, attitude, employment status, and social networks.
• Agents dynamically adjust walking attitude based on feedback mechanisms like: walking attitudes of family and friends, safety and aesthetic values, pedestrian density, and total amount walked by the person during that day.
• Agents decide walking routes and destinations based on proximity, safety, and purpose(basic needs, leisure and work).

5. Behavioral Feedback Mechanisms

1. Social Influence:

   • Attitudes update based on social connections:
     • Family:
       
     • Friends:
       

2. Walking Experience:

   • Positive walking experiences increase willingness to walk:
     

3. Density of Walkers:

   • Feedback depends on the ratio of current walking density to previous density (( I_d )):
     • ( I_d > 1 ): Positive feedback.
     • ( I_d < 1 ): Negative feedback.

4. Total Distance Walked:

   • Feedback reflects the impact of fatigue:
     

5. Daily Attitude Update:

   • Combined feedback updates attitude:
     

6. Data Collection System

• The model tracks the following metrics across five SES levels (1-5):
  • Average daily walking trips
  • Work-related trips
  • Basic needs trips (grocery and non-food shopping)
  • Leisure trips (non-purposeful neighborhood walks)

---

Known Shortcomings

1. Very slow.
2. Unable to simulate with recommended grid size (800x800) and human population.
3. Distribution of values across the safety and aesthetic property layers may not be correct.
4. Having a dog does not affect the walking attitude, whereas it should.
5. Wrong results noted:
   • 
     to be compared with
   • 
   • (one day being one step)
   • (attention to be paid to the SES color being different in both the graphs)

[EwoutH]

Looks like quite a serious model!

If you could add to the PR description:

• A summary / overview of the model
• Any outstanding questions / considerations

[EwoutH]

Thanks for the extended descriptions.

Have you profiled the model to identify the main performance bottlenecks?

[EwoutH]

@projectmesa/maintainers I’m a bit in doubt on the complexity of this example. On the one hand it shows you can create ABMs with detailed environments and complex behavior, on the other hand I don’t know an example this extended would really benefit our users. I would like to hear some of your stances.

[Sahil-Chhoker]

Have you profiled the model to identify the main performance bottlenecks?

I haven't, can you tell how can I do that?

[EwoutH]

Some starting point:

• https://docs.python.org/3/library/profile.html
• https://realpython.com/python-profiling/

Most IDEs also have a profiler built-in.

[Sahil-Chhoker]

Here is the table I built from the data collected by the profiler running the model 10 times.

Performance Analysis Results

Function/Method Name	NCalls	TotTime	CumTime	%Total
step(agents.py)	4,850	0.019	0.596	74.4%
simulate_daily_walks	4,850	0.033	0.426	53.2%
decide_leisure_walk	4,850	0.092	0.226	28.2%
calculate_distance	185,574	0.093	0.132	16.5%
get_feedback	1,303	0.041	0.128	16.0%
weakref.keys	185,980	0.072	0.113	14.1%
find_nearest_location	2,391	0.010	0.109	13.6%
min (built-in)	5,689	0.018	0.093	11.6%
sum (built-in)	14,770	0.012	0.068	8.5%
lambda (agents.py:135)	17,825	0.018	0.041	5.1%
sqrt (built-in)	190,260	0.039	0.039	4.9%
cell (cell_agent.py)	188,561	0.029	0.029	3.6%

[EwoutH]

Okay, this is quite useful already! It shows that most of the runtime is taken up in the Agent step, and specifically the decision logic.

You can try to speed that up, but if you want I can also give it a try (hopefully tomorrow or Monday).

[Sahil-Chhoker]

You can try to speed that up, but if you want I can also give it a try (hopefully tomorrow or Monday).

I'll do what I can, but I'm not particularly confident. Can we use caching to prevent duplicate lookups in calculate_distance? I'm also not sure what could speed up the other functions.

[EwoutH]

Hypothesis for model results not correlating with the paper:

1. Safety and aesthetics are note implemented correctly (check by adding visualisation of PropertyLayers)
2. Agents are not reading the correct values from the PropertyLayers
3. Agents are not using the read value correctly in their behavior

[EwoutH]

In theory this should work:

I'll help you visualize both property layers by updating the space component in app.py. We'll create a portrayal dictionary for the property layers that defines a color gradient from transparent to solid blue/green.

Here's how to modify the app.py code. First, add this portrayal dictionary right after the model_params definition:

propertylayer_portrayal = {
    "safety": {
        "color": "blue",
        "alpha": 0.5,
        "colorbar": True
    },
    "aesthetic": {
        "color": "green", 
        "alpha": 0.5,
        "colorbar": True
    }
}

Then update the space_component definition to include the property layers:

space_component = make_space_component(
    agent_portrayal, 
    draw_grid=True, 
    post_process=post_process_space,
    propertylayer_portrayal=propertylayer_portrayal
)

This will:

1. Draw the safety layer in blue gradient where low values are transparent and high values are solid blue
2. Draw the aesthetic layer in green gradient where low values are transparent and high values are solid green
3. Display colorbars for both layers to show the value scale
4. Set alpha to 0.5 so the layers blend nicely and don't completely obscure agents

The layers will be drawn before the agents, so you'll still be able to see all the agents on top of the colored background created by these layers.

[EwoutH]

• Visualize PropertyLayers mesa#2336

[Sahil-Chhoker]

I have successfully visualized the property layers, and they are functioning as intended. Before moving on to the other points, I reread the paper and realized I had overlooked the concept of households. I will now implement the household concept and check the results again.

[EwoutH]

Awesome, that's amazing to hear! Did you figure out what the bug was or what I missed yesterday?

[EwoutH]

I reread the paper and realized I had overlooked the concept of households. I will now implement the household concept and check the results again.

CC @tpike3, Households sounds like Meta agents. Maybe you could assist here!

• Meta agents mesa#2575

[Sahil-Chhoker]

@EwoutH Unfortunately no, but since they are just a bunch of numpy arrays I exported the data and made some graphs based on the values. I can confirm they are implemented as they are intended.

CC @tpike3, Households sounds like Meta agents. Maybe you could assist here!

I read a bit about MetaAgents, and from my understanding, they are groups of agents that form during the model runtime based on certain properties. While I can see their usefulness, I don’t think they are particularly relevant to this model. This is because the household system is defined during the human placement, and a household is simply the cell where they spawn. Therefore, in my opinion, this can be handled during initialization.

[EwoutH]

Thanks, sounds good!

[Sahil-Chhoker]

@EwoutH I have a question: is it correct to use model.random everywhere inside the agents' logic? For example, if agents need to find new locations to move to every day, wouldn’t using the same model.random result in the agents walking to the same place they walked the previous day?

[quaquel]

it's better to use self.random inside an Agent, so self.random.random() will draw a random number of unit interval. In either case agent.random or model.random just refer to the same seeded python stdlib random.Random instance which is a random number generator so repeated calls will return different numbers.

[Sahil-Chhoker]

The model has improved significantly with the following changes:

1. Rewriting the walking behavior to incorporate household positions.
2. Improving the placement of agents based on their SES.
3. Scaling the feedback factor by 20 due to its current size.

Tasks to be completed:

1. Fix the logic for leisure walks.
2. Review the positioning of agents.

current results for RR:

observations:

1. It becomes stable like in the graph from paper.
2. Distribution by SES is correct.

[Sahil-Chhoker]

I have thoroughly reviewed the model in every aspect and believe I have covered everything correctly. In my opinion, the only remaining task is to make the model efficient enough to run on an 800x800 grid or decide to keep this simpler model after applying smaller optimizations.

Current Results:

Current Profile: Running the model with 20x20 grid and 1350 humans for 10 steps.

Function	ncalls	tottime	percall	cumtime	percall
_wrapped_step	10	0.000	0.000	2.300	0.230
step	10	0.000	0.000	2.300	0.230
shuffle_do	10	0.023	0.002	2.165	0.217
step (agents)	13,500	0.055	0.000	2.119	0.000
simulate_daily_walks	13,500	0.101	0.000	1.636	0.000
decide_leisure_walk	13,500	0.024	0.000	0.920	0.000
get_leisure_cells	1,208	0.565	0.000	0.885	0.001
init (model)	1	0.000	0.000	0.761	0.761
add_initial_humans	1	0.007	0.007	0.715	0.715
init (agents)	1,350	0.011	0.000	0.687	0.001
get_workplace	1,350	0.245	0.000	0.561	0.000
keys	1,217,553	0.348	0.000	0.415	0.000
find_walkable_locations	8,851	0.038	0.000	0.404	0.000
get_feedback	3,001	0.133	0.000	0.369	0.000
check_near_point	10,022	0.126	0.000	0.254	0.000
sum	49,697	0.038	0.000	0.223	0.000
calculate_distance	368,985	0.137	0.000	0.212	0.000
collect	11	0.001	0.000	0.147	0.013
abs	864,370	0.125	0.000	0.125	0.000
mean	3,001	0.051	0.000	0.117	0.000
get_distance	87,277	0.083	0.000	0.103	0.000

[EwoutH]

It might be nice to add a high-level overview of the (expected) results in this Readme, and maybe a screenshot or two to support that.

[Sahil-Chhoker]

Understood!

[EwoutH]

Awesome, I will do a full review this weekend.