project_architecture.md

Project Architecture

Agents are implemented as a combination of Java and Python.

Environment

Environments or tasks should be described in Java, by extending the Environment class. Those define the observations, actions and rewards of the agent. For now, each agent has its own environment instance.

Learning Algorithms

The learning algorithms should be written in Python, they interface with Java through a MinecraftEnv instance that implements the gym interface. This environment assumes that env.reset() is called initially and after each step that returns done = True.

Communication between Java and Python

Python scripts communicate with the Minecraft server through TCP sockets.

Examples

• agent_dummy and agent_random are the most simple agent implementations
• agent_tree is an hardcoded agent for the Pattern environment
• agent_ppo and agent_acktr are wrappers for OpenAI's baselines
• agent_rl runs PyTorch RL algorithms
• agent_nearest is a K-Neareast-Neighbor agent
• test_controller is a simple usage of MinecraftController
• multi_agent_test_pattern runs multiple agents in a single environment with only one Python process
• agent_imitation_train does imitation learning