Skip to content

Creating Your Own Algorithm

This guide explains how to create your own custom multi-agent environment within the CREW Wildfire framework. In addition to our prebuilt state-of-the-art LLM-based multi-agent agentic algorithms , we offer an easy-to-use modularized template algorithm, prefitted with all the necessary functionality of a working algorithm. We offer this template algorithm with either action or primitive support, depending on your algorithm's planning frequency.

Under crew-algorithms/crew_algorithms/wildfire_alg/algorithms you should find the following structure:

algorithms
├── CAMON
├── COELA
├── Embodied
├── HMAS_2
├── MANUAL
├── TEMPLATE_no-primitives
└── TEMPLATE_primitives

No Primitives

If you do not want your agent's to use prebuilt multi-step actions, use the No Primitives template. You also would use this if your algorithm uses agents that plan a new action every timestep, making multi-step actions redundant.

Navigate to the TEMPLATE_no-primitives folder within the algorithms directory. You will find the following structure:

TEMPLATE_no-primitives
├── prompts/ 
│   ├── descriptions/ # Textual descriptions of each agent's abilities. Used in Execution Module
│   └── translator/ # Prompts for Translator/Execution Module
│
├── __main__.py # Main Script
├── agent.py # Agent Class. Contains Perception Module
└── utils.py # Utility Methods

For rapid development, we preimplemented the following infrastructure:

  • Level generation
  • Agent creation
  • Communication to the environment
  • Agent Perception from raw observations to textual descriptions (optional)
  • Agent Single-Action Execution
  • Environment Updates

The only required section to develop is the communication/planning phase, which is core to your algorithm.

This is located in the main.py file from line 202:

...
agent_actions = {}

for agent in agents:
    agent_actions[agent]=None

TODO: Implement your algorithm below:






# END Algorithm

env_action = [[0,0,0] for _ in range(cfg.envs.num_agents)]
...

After developing your algorithm's communication/planning phase, ensure that each agent's action is set in the agent_actions dictionary like so:

agent_actions[agent] = "planned action"

The action must be a string and any agent with unset actions will do nothing that timestep.

Primitives

If you want your agent to have access to prebuilt multi-step actions, use the Primitives template. You would use this if your algorithm has a dynamic planning frequency dependent on action completion.

Navigate to the TEMPLATE_primitives folder within the algorithms directory. You will find the following structure:

TEMPLATE_primitives
├── prompts/ 
│   ├── descriptions/ # Textual descriptions of each agent's abilities. Used in Execution Module
│   └── translator/ # Prompts for Translator/Execution Module
│
├── __main__.py # Main Script
├── agent.py # Agent Class. Contains Perception Module
└── utils.py # Utility Methods

For rapid development, we preimplemented the following infrastructure:

  • Level generation
  • Agent creation
  • Communication to the environment
  • Agent Perception from raw observations to textual descriptions (optional)
  • Agent Multi-Action Execution
  • Environment Updates

The only required section to develop is the communication/planning phase, which is core to your algorithm.

This is located in the main.py file from line 198:

...
for agent in agents:
    agent.generate_perception(cfg.envs, agent_states, global_data)

#TODO: Implement your algorithm below






# END Algorithm

env_action = [[0,0,0] for _ in range(cfg.envs.num_agents)]
...

After developing your algorithm's communication/planning phase, ensure that each agent's action is set in the agent_actions dictionary like so:

agent_actions[agent] = "planned action"

The action must be a string and any agent with unset actions will do nothing that timestep.

Now you can run your algorithm on created Wildfire levels.