CS330 Posterior Goal Sampling for Hierarchical Reinforcement Learning

Original Github repository : Deep Reinforcement Learning Algorithms with PyTorch.

This work was done as a class project for CS 330 : Deep Multi-Task and Meta Learning. It is a replication of Hierarchical-DQN (h-DQN), (Hierarchical Deep Reinforcement Learning: Integrating Temporal Abstraction and Intrinsic Motivation _{^{(Kulkarni et al. 2016)}} ) on the Long Corridor Game environment.

Modifications

The work was done on the h-DQN algorithm and provided a modified version that includes a Hierarchical-Ensemble DQN RL for efficient high-level policy learning. The code is present in agents/hierarchical_agents/bh_agents.py.

We also included a new function pick_actions2() in agents/DQN_agents/DQN.py that allows goal sampling among all Q heads functions.

Execution

python results/Long_Corridor.py

Usage

The repository’s high-level structure is:

├── agents                    
    ├── DQN_agents         
    ├── actor_critic_agents   
    ├── hierarchical_agents
    └── policy_gradient_agents
├── environments   
├── exploration_strategies
├── results             
    └── data_and_graphs        
├── tests
├── utilities             
    └── data structures