RandomActivation scheduler not working after upgrade 2.1 -> 2.2.3

[profConradi]

I upgraded 2.1 to 2.2.3 and in this simple test model RandomActivation scheduler does not work anymore: agent activation is done sequentially.
Furthermore if I not call super().init() in the model class, the first agent is not added to the scheduler. Is this a bug or a mistake on my part?

    class CognitiveAgent(mesa.Agent):
        """Agents make decisions using automatic (e.g., reflexive) versus controlled (e.g., deliberative) cognition,
        interact with each other, and influence the environment (i.e., game payoffs)."""
    
        def __init__(self, unique_id, model):
            # Pass the parameters to the parent class.
            super().__init__(unique_id, model)
    
            # Each agent is defined by a single parameter x that represents its probability of controlled processing
            self.xdx = np.random.uniform(0., 1.)
    
        def step(self):
            # The agent's step will go here.
            # For demonstration purposes we will print the agent's unique_id
            other_agent = self.random.choice(self.model.schedule.agents)
            print(f"I'm {self.unique_id} and I choose {other_agent.unique_id}")
            
    class CognitiveModel(mesa.Model):
        """A model with some number of agents."""
    
        def __init__(self, N):
            super().__init__()
            self.num_agents = N
            self.schedule = mesa.time.RandomActivation(self)
            # Create agents
            for i in range(self.num_agents):
                a = CognitiveAgent(i, self)
                # Add the agent to the scheduler
                print(a,i)
                self.schedule.add(a)
    
        def step(self):
            """Advance the model by one step."""
            # The model's step will go here for now this will call the step method of each agent and print the agent's unique_id
            self.schedule.step() 

[EwoutH]

Thanks for reaching out. I will test this tomorrow morning, if no one else has done it by then.

Furthermore if I not call super().init() in the model class, the first agent is not added to the scheduler.

Calling super().__init__() was always recommended, but now it’s required. It initializes Model.agents among other things. You’re using the Mesa Model class, so you should properly initialize it.

[EwoutH]

@rht do you have time to investigate this?

[rht]

I'd need some help debugging all the problems caused by 2.1.x -> 2.2.x. Hands are full with elsewhere.

[EwoutH]

Unfortunately, I can reproduce this issue for both Mesa 2.2.0 and 2.2.3, while the correct behavior (shuffled agent steps) on Mesa 2.1.5. @projectmesa/maintainers this is a major issue and needs immediate care, we broke RandomActivation.

I will investigate further.

[EwoutH]

I've narrowed it down to the fact that the AgentSet's shuffle(inplace=True) doesn't shuffle the agents. Could be because the shuffle isn't working, or the in place updating isn't.

mesa/mesa/time.py

Lines 144 to 147 in 99b35fc

	def do_each(self, method, shuffle=False):
	if shuffle:
	self.agents.shuffle(inplace=True)
	self.agents.do(method)

[EwoutH]

It's the inplace updating that isn't working. Without it, it works correctly:

@quaquel you wrote this code and know more about the AgentSet implementation. Could assist in fixing the inplace updating?

[profConradi]

Thanks for your work, Mesa is an amazing framework!

[quaquel]

I can try this afternoon. I don't see an obvious mistake in the code.

AgentSet.shuffle

        shuffled_agents = list(self)
        self.random.shuffle(shuffled_agents)

        return (
            AgentSet(shuffled_agents, self.model)
            if not inplace
            else self._update(shuffled_agents)
        )

random.shuffle is inplace. List(self) returns a list of hard refs to the agents.

AgentSet._update

    def _update(self, agents: Iterable[Agent]):
        """Update the AgentSet with a new set of agents.
        This is a private method primarily used internally by other methods like select, shuffle, and sort.
        """

        self._agents = weakref.WeakKeyDictionary({agent: None for agent in agents})
        return self

We create a new WeakKeyDictionary so it should shuffle.

Accidentally, This is the code I was playing with for #1993. So my to do;

• add a unittest that reproduces the error
• figure out why the error exists, because based on the code I don't
• test my fix from replace weakref dict with normal dict #1993

[EwoutH]

Thanks. I had already written an additional unittest for RandomActivation itself, we can test updates also against that (#2007).