khanhvu207 / deer-pytorch

PyTorch implementation of Combined Reinforcement Learning via Abstract Representations

PyTorch implementation of Combined Reinforcement Learning via Abstract Representations

Khanh Vu

Vrije Universiteit Amsterdam

Introduction

During my Bachelor's years 2021-2022, I was fortunated to work with Prof. Dr. Vincent François-Lavet on a reinforcement learning project with a focus on representation learning. This repository features my PyTorch implementation of his paper Combined Reinforcement Learning via Abstract Representations. The core algorithm consists of a contrastive representation learning module for the unsupervised learning of world models and a model-free Q-learning for decision making and planning. In the project, I worked on the world models learning part specifically.

Installation

1. Create a virtual environment

python3 -m venv venv
source venv/bin/activate

2. Install dependencies

pip3 install -r requirements.txt

Run the model

1. Modify the config.yaml file to change the learning environment and hyperparameters of the model.

experiment_args:
  steps_per_epoch: 5000
  epochs: 10
  steps_per_test: 5000

env_args:
  env_name: "loop_maze_full"
  size_x: 5
  size_y: 5
  higher_dim_obs: False

train_args:
  internal_dim: 2
  update_rule: 'rmsprop'
  learning_rate: 0.0005
  learning_rate_decay: 0.5
  rms_decay: 0.9
  rms_epsilon: 0.0001
  momentum: 0
  clip_norm: 1.0
  beta2: 0.1
  C: 2
  radius: 1.0
  epsilon_start: 1.0
  epsilon_min: 1.0
  epsilon_decay: 10000
  update_frequency: 1
  replay_memory_size: 1000000
  batch_size: 32
  freeze_interval: 1000
  deterministic: True
  discount: 0.9
  discount_inc: 1
  discount_max: 0.99
  device: 'cpu'

logger_args:
  online_mode: False

2. Run the model via

python3 main.py