Experiment - Sugarscape

[bordumb]

Plan for Implementing an ARLA-Powered Sugarscape

This plan outlines a two-phase approach to building a Sugarscape-style simulation. Phase 1 focuses on faithfully replicating the core mechanics described in the paper to establish a validated baseline. Phase 2 introduces advanced cognitive features unique to the ARLA platform to explore research questions beyond the original study.

Phase 1: Baseline Replication

The goal of this phase is to recreate the fundamental environment and agent behaviors from the paper. This will allow us to validate our implementation by reproducing the emergent phenomena of cooperation and competition.

1. Environment Implementation (SugarscapeEnvironment)

We will adapt the existing BerryWorldEnvironment to create a new SugarscapeEnvironment. The core changes will be:

Resource Model: Replace the discrete berry locations (berry_locations) with a continuous sugar level for each grid cell. We'll add a sugar_map attribute, a 2D NumPy array representing the sugar concentration at each (x, y) coordinate.

Regeneration: Implement a regenerate_sugar() method that is called each tick. This method will increment the sugar level of all cells up to a predefined maximum, simulating resource renewal.

Agent Metabolism: The environment will not directly manage agent energy. Instead, it will provide methods like get_sugar_at(position) and consume_sugar(position, amount) that systems can call.

2. New and Modified Components

We'll need to introduce new components to represent the state of a Sugarscape agent.

EnergyComponent (New): A simple component to store the agent's current energy level. This replaces the HealthComponent.

MetabolismComponent (New): Stores the agent's metabolic rate (energy consumed per tick) and vision range. This allows for agent heterogeneity, a key feature of the original Sugarscape.

PositionComponent (Existing): No changes needed.

CommunicationComponent (New, from your simple_sim): We will add a component to handle messaging, similar to the paper's description of agents communicating within a 7x7 range.

3. Core Agent Actions

We will implement the actions described in the paper, registering them with the action_registry.

MoveAction: Will now have a variable energy cost based on the agent's MetabolismComponent.

HarvestAction (New): An action for gathering sugar from the current cell.

ShareAction (New): Allows an agent to transfer energy to another agent.

AttackAction (New): Allows an agent to attack another agent to steal its energy.

ReproduceAction (New): Creates a new agent, splitting the parent's energy.

4. World-Specific Systems

These systems will contain the logic for how actions affect the world and the agents.

MetabolismSystem (New): At the start of each tick, this system will iterate through all agents with an EnergyComponent and MetabolismComponent and deduct the energy cost for that tick. If an agent's energy reaches zero, this system will deactivate it.

HarvestSystem (New): Subscribes to the execute_harvest_action event. When an agent harvests, this system will call env.consume_sugar() and add the energy to the agent's EnergyComponent.

SocialSystem (New): A single system that subscribes to execute_share_action, execute_attack_action, and execute_reproduce_action to handle the logic for these complex social interactions.

Phase 2: Extending with ARLA's Cognitive Architecture

With a validated baseline, we can now introduce ARLA's unique cognitive features to ask more nuanced research questions.

1. Introducing Subjective Experience (AffectSystem)

The paper observes emergent strategies but doesn't model the agent's internal experience. We can introduce the AffectSystem to explore how emotion influences behavior.

Hypothesis: Agents in a "fearful" emotional state (low valence, high arousal) will be less likely to attack, even in scarce conditions, compared to agents in an "angry" state.

Implementation:

• Add EmotionComponent and AffectComponent to the agents
• Create a SugarscapeVitalityProvider that maps high energy to positive valence and low energy to negative valence
• The QLearningDecisionSelector will now receive an internal_state_vector that includes the agent's current emotional state, allowing the policy to learn from emotions

2. Moving from Heuristics to True Causal Reasoning (CausalGraphSystem)

The paper's agents use LLM reasoning, but it's based on correlation. We can test if agents with a formal causal model develop more robust strategies.

Hypothesis: A causal agent will learn that attacking leads to a net energy gain only when its own energy is low, whereas a standard RL agent might incorrectly learn that attacking is always a good strategy after a few initial successes.

Implementation:

• Enable the CausalGraphSystem
• Create a SugarscapeStateNodeEncoder that encodes the agent's state into abstract concepts like "energy_level": "critical" and "local_sugar": "abundant"
• The QLearningSystem will then use the output of the agent's learned causal model to get a more robust reward signal, blending the observed reward with the causal estimate

3. Complex Social Dynamics (IdentitySystem & SocialMemoryComponent)

We can model the emergence of social structures like tribes or alliances by enabling agents to form identities and remember their interactions.

Hypothesis: Agents will be more likely to share with other agents they identify as "kin" (from reproduction) or "allies" (from repeated positive interactions), even under moderate scarcity.

Implementation:

• Add IdentityComponent and SocialMemoryComponent to the agents
• The SocialSystem will be updated to log interactions (attacks, shares) in the SocialMemoryComponent
• The ShareAction's generate_possible_params method will be modified to be more likely to generate share actions directed at agents with a positive relationship score

By following this plan, you will not only replicate the fascinating results from the paper but also push the boundaries of the research by exploring the impact of emotion, causal reasoning, and social identity on emergent survival strategies.

[bordumb]

Source: https://arxiv.org/pdf/2508.12920