Mesa mixes up agent objects and runs another agent's step method

[jingxuanlim]

Hi. I ran into an issue that I feel like is a bug, because it works fine for multiple steps then stops working after.

Assume that I have two agent classes, Predator and Prey and they have different attributes and step methods. At this point, the two classes don't interact -- they just move. I found that they move fine for a variable number of steps, and then runs into the error (find output below, which ran for 10 steps but other times it would run for 6 or 1 etc).

The traceback states that Predator is now trying to run Prey's step method which relies on an attribute that only the Prey has (velocity in this case).

It therefore seems like mesa mixed up the two agent objects' methods, leading to an error. Has anyone experienced this?

It's possible that there's a problem with my code, which I am not able to find. Any hints welcomed!

WARNING:tornado.access:404 GET /favicon.ico (127.0.0.1) 3.50ms
{"type":"get_step","step":1}
{"type":"get_step","step":2}
{"type":"get_step","step":3}
{"type":"get_step","step":4}
{"type":"get_step","step":5}
{"type":"get_step","step":6}
{"type":"get_step","step":7}
{"type":"get_step","step":8}
{"type":"get_step","step":9}
{"type":"get_step","step":10}
ERROR:tornado.application:Uncaught exception GET /ws (127.0.0.1)
HTTPServerRequest(protocol='http', host='127.0.0.1:8521', method='GET', uri='/ws', version='HTTP/1.1', remote_ip='127.0.0.1')
Traceback (most recent call last):
.
.
.
AttributeError: 'Predator' object has no attribute 'velocity'

[rht]

If you have a GitHub repo I can checkout, it will be much easier to debug.
Edit: answer can be found in the thread below.

[rht]

I found the cause. If the model is private, I can tell you at our Matrix.

[jingxuanlim]

Ah, no problem. I barely got started on it. Discussing in this thread is fine. Appreciate the thought though!

[rht]

The reason is because the Prey inherits from Boid. Boid has this method

    def match_heading(self, neighbors):
        """
        Return a vector of the neighbors' average heading.
        """
        match_vector = np.zeros(2)
        if neighbors:
            for neighbor in neighbors:
                match_vector += neighbor.velocity
            match_vector /= len(neighbors)
        return match_vector

i.e. it doesn't check to make sure that the neighbor is a Prey. You have to override the match_heading method in the Prey class declaration, so that it checks whether the neighbor has velocity as an attribute.

[rht]

Actually, I think it's better to check by the class of the neighbors isinstance(neighbor, Prey), rather than checking whether a neighbor has a velocity attribute.

[jingxuanlim]

Ahhh, silly me...! Of course I was wrong to assume that the agents are not interacting with each other! Works as expected now!

Also amazing that you were able to debug so quickly, despite having to look at someone else's code! Very impressed! Thanks!!

[jingxuanlim]

@rht One more thing I thought I ask here to get a quick answer. I hope you don't mind.

How do I get a video recording of the model running? Besides doing a screen recording, is there a way to do it from mesa?

[rht]

Hmm, never had to do that. I assume you want it like https://asciinema.org/ does for terminal sessions, where the recording output is semantic and much smaller than a video. Not sure this is the right path, but you might try extracting the output of the Chrome profiler (on Chrome/Chromium/Brave). You press ctrl-shift-I, and then go to the performance tab, and then you record. You should be able to see the rendered page being replayable.

[rht]

Agents.jl surely does have a nice feature having easily recordable animation: see https://juliadynamics.github.io/Agents.jl/v4.1/examples/sugarscape/. We might be able to do this using Matplotlib. But not sure how to do this given with custom agent portrayals that are not just circles.