Generative Agents Collaboratively Mission-Planning

This repository explores emergent social behaviors in generative agents, building upon the paper "Generative Agents: Interactive Simulacra of Human Behavior." The project specifically focuses on simulating party environments to study complex social dynamics and emergent behaviors.

---

Index:

1. Project Overview
2. Setup
3. Execution
4. Cost-Tracking
5. Customizing the Map
6. Party Simulation

---

Project Overview

This project investigates how autonomous agents form social groups, share information, and adapt their plans within structured social environments. Key research areas include:

• Formation of social groups based on memory and perception
• Information diffusion in locally-constrained environments
• Impact of limited memory retrieval on decision-making
• Real-time plan adaptation
• Emergent patterns in dynamic schedule adjustment

Simulation Types

• Informal Get-together: Casual social interactions
• Formal Networking Event: Professional networking dynamics

Quantifiable Metrics

• Information Spread (with whisper function)
• Social Clustering
• Adaptation Rate
• Interaction Density
• Acceptance/Rejection Rate
• Agent Mobility
• Conversation Duration

Visualizing Metrics

The project includes a comprehensive visualization system to analyze emergent behaviors in the simulation. To generate visualizations:

1. Install the required dependencies:

pip install matplotlib seaborn networkx pandas numpy

2. Run the visualization script:

python reverie/backend_server/persona/insight/visualize_metrics.py

This will generate several visualizations in the visualizations directory:

• Interaction Network: Shows the social network structure of agents with edge thickness representing interaction frequency
• Interaction Heatmap: Displays the frequency of interactions between all pairs of agents
• Information Spread: Visualizes how much information each agent has received
• Acceptance/Rejection Rates: Shows the ratio of successful vs failed interactions
• Zone Movements: Displays agent mobility patterns between different zones
• Conversation Durations: Shows the distribution of conversation lengths

These visualizations help analyze:

• Social network formation and dynamics
• Information propagation patterns
• Success rates of social interactions
• Spatial behavior and movement patterns
• Typical conversation characteristics

Party Simulation

Environment Structure

The party environment is modeled as a tree-structured space with distinct zones:

• Bar area
• Dance floor
• Lounge
• Entrance area

Agent Roles

1. Host (1 agent)

   • Acts as a daemon influencing all agents
   • Rules and personality defined by the group

2. Guests (4 agents)

   • 3 agents modeled as digital twins of group members
   • 1 agent with randomized traits

Agent Capabilities

• Local perception within their zone
• Memory stream storage (observations, reflections, plans)
• Dynamic plan modification
• Information sharing through whisper function
• Team formation and coordination

Setting Up The Environment

Generate environment

To created the image in your environment run this code.

   python tools/generate_env_images.py

Step 1. Conda Env

Note: If you change the environment name from simulacra, you'll need to update the name in the upcoming bash scripts as well.

    conda create -n simulacra python=3.9.12 pip
    conda activate simulacra
    pip install -r requirements.txt

Download Famous People Images

To download images of famous people for testing purposes, run:

python tools/download_famous_people_images.py

Run Simulations

All of the following scripts accept two optional arguments to customize the conda setup:

• --conda_path: Path to your conda activate script (default: /home/${USER}/anaconda3/bin/activate)
• --env_name: Name of the conda environment to use (default: simulacra)

Example with custom conda setup:

./run_frontend.sh --conda_path /opt/anaconda3/bin/activate --env_name simulacra
./run_frontend.sh --env_name simulacra

Example:

   ./run_backend_automatic.sh --conda_path /opt/anaconda3/bin/activate --env_name simulacra -o base_party -t test_1 -s 400 --ui True

   # Example without UI
   ./run_backend_automatic.sh --env_name simulacra -o base_party -t party_experiment_1_whisper_2 -s 800 --ui True
   ./run_backend_automatic.sh --env_name simulacra -o base_party -t party_experiment_1_whisper_5 -s 800 --ui True
   ./run_backend_automatic.sh --env_name simulacra -o base_party -t medical_experiment_1_whisper_2 -s 800 --ui True
   ./run_backend_automatic.sh --env_name simulacra -o base_party -t medical_experiment_1_whisper_5 -s 800 --ui True

   # In case you want to continue a simulation, search for the latest simulation
   ./run_backend_automatic.sh --env_name simulacra -o party_experiment_1_whisper_1-s-3-599-799 -t party_experiment_1_whisper_1 -s 800 --ui True

   ./run_backend_automatic.sh --env_name simulacra -o party_experiment_1_whisper_2-s-3-599-799 -t party_experiment_1_whisper_2 -s 800 --ui True

   # Server 
   ./run_backend_automatic.sh --conda_path /home/gonzalaa/.miniconda3/bin/activate --env_name simulacra -o base_party -t party_experiment_1_whisper_50 -s 800 --ui None

   ./run_backend_automatic.sh --env_name simulacra -o medical_experiment_1_whisper_2-s-2-399-599 -t medical_experiment_1_whisper_2 -s 800 --ui True

   ./run_backend_automatic.sh --env_name simulacra -o party_experiment_1_whisper_5-s-1-199-399 -t party_experiment_1_whisper_5 -s 800 --ui True

   ./run_backend_automatic.sh --env_name simulacra -o base_party -t final_validation -s 100 --ui True
      
   ./run_backend_automatic.sh --conda_path /home/gonzalaa/.miniconda3/bin/activate --env_name simulacra -o base_party -t party_experiment_1_whisper_50 -s 800 --ui None
   
   # New test base on latest suggestions
   ./run_backend_automatic.sh --env_name simulacra -o base_party -t party_experiment_1_whisper_50 -s 800 --ui True

   ./run_backend_automatic.sh --env_name simulacra -o base_party -t party_experiment_os_1_whisper_50 -s 800 --ui True

   ./run_backend_automatic.sh --env_name simulacra -o base_party -t party_experiment_os_1_whisper_50_ilegal -s 800 --ui True


   # With past experiences
   ./run_backend_automatic.sh --env_name simulacra -o base_party -t crossmodality_plan_short7 -s 60 --ui True

   ./run_backend_automatic.sh --env_name simulacra -o base_party_memory -t new_crossmodality_0 -s 2400 --ui True

   ./run_backend_automatic.sh --env_name simulacra -o base_party -t new_crossmodality_0 -s 600 --ui True --scenario_index 5

   # 1800 + 600

Endpoint list

• http://localhost:8000/ - check if the server is running
• http://localhost:8000/simulator_home - watch the live simulation
• http://localhost:8000/replay/<simulation-name>/<starting-time-step> - replay a simulation

For a more detailed explanation see the original readme.

Running a Demo

To create a polished demo with proper character sprites:

1. First, compress your simulation:

   • Open reverie/compress_sim_storage.py
   • Run the compress() function with your simulation name

   compress("your_simulation_name")

2. View the demo in your browser:

   http://localhost:8000/demo/<simulation-name>/<starting-time-step>/<simulation-speed>
   

   Parameters:

   • simulation-name: Name of your simulation
   • starting-time-step: Where to begin the demo (e.g., 1)
   • simulation-speed: Speed from 1 (slowest) to 5 (fastest)

Example demo URL: http://localhost:8000/demo/test_1-s-0-0-186_good/1/3/

Authors

This project was developed as part of the Complex Systems course under the guidance of Professor Ali A. Minai.

Team Members

• Adonai Vera
• Carlos Hinojosa
• Karen Sanchez
• Prateek Kharangate
• Haidar Bin Hamid