Fix: Schelling Model Neighbor Similarity Calculation

[Sahil-Chhoker]

Summary

Fixed #2515 Schelling segregation model to calculate agent happiness using neighbor similarity fraction, aligning with Wikipedia and NetLogo standards.

Bug / Issue

Current model incorrectly counts empty spaces as neighbors, leading to inaccurate agent happiness determination.

Implementation

Modified SchellingAgent.step() to:

• Filter out empty cells
• Calculate similar neighbor fraction
• Scale homophily threshold

similar_neighbors = [
    neighbor for neighbor in neighbors 
    if hasattr(neighbor, 'type') and neighbor.type == self.type
]
total_neighbors = [
    neighbor for neighbor in neighbors 
    if hasattr(neighbor, 'type')
]

if len(total_neighbors) > 0:
    similarity_fraction = len(similar_neighbors) / len(total_neighbors)
    
    if similarity_fraction < self.model.homophily / 8.0:
        self.model.grid.move_to_empty(self)
    else:
        self.model.happy += 1

[github-actions]

Performance benchmarks:

Model	Size	Init time [95% CI]	Run time [95% CI]
BoltzmannWealth	small	🔴 +5.4% [+3.5%, +7.2%]	🔵 +0.4% [+0.3%, +0.5%]
BoltzmannWealth	large	🔵 -0.3% [-0.9%, +0.2%]	🔵 -3.5% [-6.0%, -1.0%]
Schelling	small	🔵 +0.6% [+0.3%, +0.9%]	🟢 -16.1% [-16.3%, -15.8%]
Schelling	large	🔵 +0.4% [-0.0%, +0.9%]	🟢 -6.5% [-7.6%, -5.4%]
WolfSheep	small	🔵 +0.2% [-0.1%, +0.5%]	🔵 -0.0% [-0.2%, +0.1%]
WolfSheep	large	🔵 -1.4% [-2.3%, -0.5%]	🔵 -3.2% [-5.9%, +0.1%]
BoidFlockers	small	🔵 +0.9% [+0.4%, +1.5%]	🔵 +0.7% [-0.2%, +1.7%]
BoidFlockers	large	🔵 +0.4% [-0.4%, +1.2%]	🔵 +0.1% [-0.3%, +0.5%]

[quaquel]

why not change homopily at the model level to be a fraction? That makes the model independent of the neighborhood size.

[quaquel]

you iterate twice, with one check being done in both. You can make this more efficient by looping only once.

[EwoutH]

@Sahil-Chhoker Thanks for your PR. When do you expect to be able to incorporate @quaquel's feedback?

If you have any question about it feel free to ask!

[Sahil-Chhoker]

Thank you, @quaquel, for your review! Could you please provide more details about the model and how can I can make changes to it?

[quaquel]

My feedback is quite clear, and there is a link to a more detailed description of the model in the original issue. I requested 2 well-defined changes to this PR: change homophily to a fraction and rewrite so you only loop once.

From your reaction, I therefore deduce that you are not particularly familiar with agent-based modeling.

[quaquel]

this is not needed, iter_neighbors returns a list of agents. In this model all agents have the type attribute.

[quaquel]

        neighbors = self.model.grid.get_neighbors(
            self.pos, moore=True, radius=self.model.radius
        )

        # Count similar neighbors
        similar_neighbors = len([n for n in neighbors if n.type == self.type])

        # Calculate the fraction of similar neighbors
        if (valid_neighbors := len(neighbors) )> 0:
            similarity_fraction = similar_neighbors / valid_neighbors

            # If unhappy, move to a random empty cell
            if similarity_fraction < self.model.homophily:
                self.model.grid.move_to_empty(self)
            else:
                self.model.happy += 1

[quaquel]

see my second round of feedback

[quaquel]

Please switch from iter_neighbors to get_neighbors.

        neighbors = self.model.grid.get_neighbors(
            self.pos, moore=True, radius=self.model.radius
        )

currently, as you can see the tests fail because the model won't run. The list expression exhausts the iterator, so getting the number of valid neighbors fails.

[github-actions]

Performance benchmarks:

Model	Size	Init time [95% CI]	Run time [95% CI]
BoltzmannWealth	small	🔵 -1.5% [-2.1%, -0.9%]	🔵 -1.6% [-1.8%, -1.3%]
BoltzmannWealth	large	🔵 -2.8% [-5.8%, -1.1%]	🔵 -2.8% [-3.4%, -2.4%]
Schelling	small	🔵 -1.9% [-2.4%, -1.5%]	🔴 +82.8% [+81.9%, +83.7%]
Schelling	large	🔵 -1.5% [-1.9%, -1.2%]	🔴 +85.2% [+83.9%, +86.6%]
WolfSheep	small	🔵 -0.3% [-0.5%, -0.1%]	🔵 -0.5% [-0.7%, -0.2%]
WolfSheep	large	🔵 +0.6% [-0.1%, +1.4%]	🔵 +5.2% [+2.4%, +8.2%]
BoidFlockers	small	🔵 -0.6% [-1.1%, -0.0%]	🔵 -0.2% [-1.0%, +0.5%]
BoidFlockers	large	🔵 -0.6% [-0.9%, -0.3%]	🔵 -0.0% [-0.6%, +0.4%]

[quaquel]

what happens if len(neighbors) is 0? Currently, nothing happens, which most definitely is not correct.

[Sahil-Chhoker]

Please correct me if I’m wrong, but if there are no neighbors surrounding the agent, it should be considered unhappy as it does not meet the criteria for happiness. Should I proceed with this approach?

[quaquel]

I have checked some literature, including the original article. It's not entirely clear. However, your reading is defendable, so I am fine with it.

[github-actions]

Performance benchmarks:

Model	Size	Init time [95% CI]	Run time [95% CI]
BoltzmannWealth	small	🔵 +0.1% [-0.7%, +0.9%]	🔵 +0.3% [-0.4%, +0.9%]
BoltzmannWealth	large	🔵 +0.8% [-0.5%, +2.2%]	🔵 +0.7% [-0.3%, +1.6%]
Schelling	small	🔵 +0.1% [-0.6%, +0.9%]	🔴 +87.7% [+86.7%, +88.7%]
Schelling	large	🔵 +0.0% [-0.3%, +0.3%]	🔴 +84.7% [+83.9%, +85.5%]
WolfSheep	small	🟢 -4.1% [-4.8%, -3.3%]	🟢 -8.9% [-9.3%, -8.4%]
WolfSheep	large	🟢 -3.7% [-4.0%, -3.4%]	🟢 -10.4% [-10.9%, -9.8%]
BoidFlockers	small	🔵 +0.5% [-0.4%, +1.5%]	🔵 -1.3% [-2.0%, -0.6%]
BoidFlockers	large	🔵 +0.7% [-0.1%, +1.7%]	🔵 -1.2% [-1.5%, -0.8%]

[EwoutH]

Thanks for bringing this PR this far. Is the performance regression expected from the code changes, or is it bigger than expected?

[quaquel]

I am not surprised as indicated before, but I still want to check before merging.

[quaquel]

I have checked the model, and something is not quite right. Within a few steps, all agents are supposedly happy but if you check the grid manually, you see that this cannot be true. I am not yet sure what's causing this. It might be merely the order of agent activation (so an agent goes first and is happy, but some of its neighbors move afterwards, meaning it is unhappy if checked again). I need time to look more closely at this and reason true what is going on.