datenwissenschaften 2.5.1

Core library for Retro Speedlab: A high-performance RL toolkit for classic games.

datenwissenschaften

The reinforcement-learning engine behind Retro Speedlab.

datenwissenschaften turns classic-game emulators into reproducible training systems. It provides visual and state-aware environments, recurrent PPO agents, parallel execution, durable checkpoints, episode recording, and a live browser dashboard in one focused Python package.

This repository contains the reusable engine. For game runners, end-to-end examples, and user-facing documentation, start with Retro Speedlab.

Why this engine

• Exploration for sparse rewards — multi-input CNN-LSTM PPO combines visual frames, normalized RAM, temporal memory, and normalized, clipped, annealed Random Network Distillation (RND).
• Efficient execution — vectorized environments, automatic worker selection, CUDA tuning, and CPU fallback.
• Reliable training runs — atomic checkpoints, resumable model state, automatic savestates, and .bk2 replay capture.
• Operational visibility — a local Vue dashboard reports episode outcomes, reward distributions, environment details, PPO parameters, RND progress, and automatic-savestate progress.
• Game-oriented infrastructure — ROM discovery, RAM models, state machines, visual encoders, and configurable action translation.

Model choices

Model	Best suited to	Characteristics
RecurrentRNDModel	Sparse-reward NES games and partially observable state	Automatic visual + RAM inputs, NES-tuned recurrent PPO, LSTM memory, intrinsic RND exploration
Custom SB3 model	Experiments that need a standard Stable-Baselines3 algorithm	Integrates through the same builder, trainer, callbacks, and dashboard

RecurrentRNDModel is the recommended starting point for visual agents. RND encourages the policy to visit novel observations, while its influence decays during training so learned external rewards increasingly drive behavior. The predictor, fixed target, optimizer, reward statistics, and annealing progress are all preserved in checkpoints. Environment wrappers always use RGB observations and one emulator step per selected action; these are fixed engine defaults rather than game-level options. The default profile uses longer 512-step rollouts, a 512-unit LSTM, gamma=0.999, gae_lambda=0.98, and a slower 10-million-step RND decay. These settings preserve more temporal context and delayed reward information than the shorter arcade baseline while retaining conservative PPO updates.

Set training.savestate_beaten_threshold to the number of victories required before a training state is marked as beaten and the next automatic savestate is promoted. Each <State>.beaten file stores the current victory count; the default threshold is 1. PPO telemetry and automatic-savestate safeguards use extrinsic environment reward only; RND curiosity remains an exploration bonus for policy updates and does not count as savestate progress. When an automatic savestate is written, the wrapper stores the cumulative extrinsic reward reached before that state in a matching <State>.reward file. Episodes restored from that savestate receive the saved reward as a one-time baseline, then state rewards and penalties add to or subtract from that baseline.

Installation

The package requires Python 3.12.

pip install datenwissenschaften

For local development:

git clone https://github.com/datenwissenschaften/datenwissenschaften.git
cd datenwissenschaften
poetry install
cp config.example.yaml config.yaml

Training dashboard

Enable the dashboard with ui.enable: true, then open http://127.0.0.1:18080. It refreshes live training telemetry without interrupting the learner.

Dashboard history is restored from and persisted to Redis. The default Redis URL is redis://127.0.0.1:6379/0, and history keys use the datenwissenschaften:history prefix. The ui mapping accepts enable, host, port, max_episodes, redis_url, and history_key_prefix. Snapshots retain the latest 1,000 episodes by default and include summarized totals for discarded episodes; set max_episodes to another positive integer or null for unlimited retained rows. Binding to 0.0.0.0 makes the dashboard reachable on the local network; use that only on a trusted network and open it through the machine's actual IP address.

How a run fits together

1. A game package defines RAM structures, training states, rewards, and action translation.
2. The environment factory creates vectorized emulator workers and processed visual observations.
3. A model builder creates or restores the selected policy.
4. The trainer coordinates learning, checkpoints, replay capture, telemetry, and optional uploads.
5. The dashboard exposes the active run without coupling the learner to a separate monitoring service.

Development

Run Python quality checks:

ruff check src
black --check src
python -m compileall -q src

Build the dashboard assets after changing the Vue frontend:

cd src/datenwissenschaften/ui/frontend
npm ci
npm run build

License

Copyright © datenwissenschaften contributors. Distributed under the GNU General Public License v3.0.