CatkinChen/nethack-ppo-ablation-baseline

This repository contains a complete Sequential Skill RL model trained on NetHack/MiniHack environments.

Model Components

1. PPO Policy (ppo_policy.pth)

• Type: Proximal Policy Optimization agent
• Environment: MiniHack-Room-5x5-v0
• Training Steps: 50,000
• Features:
  • Curiosity-driven exploration: True
  • Random Network Distillation: False

2. VAE Model (vae_model.pth)

• Type: Variational Autoencoder
• Purpose: Encodes NetHack observations into latent skill representations
• Latent Dimension: 96
• Architecture: Unknown

3. HMM Model (hmm_model.pth)

• Type: Hidden Markov Model (Sticky HDP-HMM)
• Purpose: Models sequential skill transitions and dynamics
• Integration: Used for intrinsic reward computation

Usage

import torch
from training.online_rl import train_online_ppo_with_pretrained_models

# Load the complete model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Load individual components
ppo_checkpoint = torch.load('ppo_policy.pth', map_location=device)
vae_data = torch.load('vae_model.pth', map_location=device)
hmm_data = torch.load('hmm_model.pth', map_location=device)

# Use for inference or continued training
results = train_online_ppo_with_pretrained_models(
    env_name="MiniHack-Room-5x5-v0",
    vae_repo_id="CatkinChen/nethack-vae-hmm",
    hmm_repo_id="CatkinChen/nethack-hmm",
    test_mode=True
)

Training Configuration

• Environment: MiniHack-Room-5x5-v0
• Learning Rate: 0.0005
• Batch Size: 32
• Training Time: 0.02 seconds
• Device: cuda
• Seed: None

Performance

Training completed successfully with the following configuration:

• Curiosity-driven exploration: True
• Random Network Distillation: False

Generated on: 2025-09-19 14:38:50