Imitation and Reinforcement Learning in Minecraft

Context

This project was developed during my 5 months academic internship at USC, Los Angeles. Under the supervision of Fei Sha, my task was to develop a learning framework within Minecraft. This framework was aimed at Imitation Learning and Reinforcement Learning. I also experimented with opportunities for learning task using Minecraft.

The long term goal is to use Minecraft to experiment with learning in the context of human societal behaviour.

A new Framework

The resulting framework has many features. It can help with Imitation Learning by providing convenient solutions to record gameplay. Reinforcement Learning environments can easily be implemented and some input and output spaced are already defined. The main feature is to allow usage of Python agents and fast game execution to speed up learning.

    alt
Simple raytracing input space with one color per block type
    alt
Abstract raytracing for a team game environment

Scientific Report

More details can be found in my detailed Scientific Report.

    alt
Convolutional Neural Network Architecture

Animated Figures

Animated Figure 1: Comparison between direct screen recording and screen recording after reconstruction according to our recording format
Animated Figure 2: Parkour environment after 9 million time steps of A2C
Animated Figure 3: CowArena environment with attention layer in the training setting (1-2 cows)
Animated Figure 4: CowArena environment with attention layer transfer to another setting (1-20 cows)
Animated Figure 5: Simple break tree policy: data
Animated Figure 6: Simple break tree policy: imitation
Animated Figure 7: Human subjects: data
Animated Figure 8: Human subjects: imitation
Marc Ducret copyright © 2024