Training
Configure experiments, launch reproducible runs, and monitor progress in real time. Every run is versioned and traceable.
Experiment configuration
Create experiments and define your RL setup before you run. All configuration is explicit and versioned.
- Environment: symbols, timeframe, and market data from your broker
- Policy: algorithm and hyperparameters (e.g. PPO)
- Actions, observations, and reward: full control over the RL loop
- Validation: schema and server-side checks before execution
Reproducible runs
Launch training runs tied to an experiment. Config and execution are versioned so you can reproduce results every time.
Each run is linked to an experiment and records the configuration at launch time. Resume from checkpoints when you need to extend training or recover from interruptions.
Real-time monitoring
Watch training progress, key metrics, and failures live from the dashboard. Updates stream over realtime; the UI falls back to a short poll only if the stream is unavailable.
Progress and metrics are pushed over a realtime connection so you see updates as soon as they’re available. Track episodes, rewards, and failures without manual refresh. If the stream drops, the dashboard briefly polls so you are not stuck on a blank board.
Traceable artifacts
Every run logs metrics and artifacts. Review run history, manage models, and keep experiments auditable.
Models and metrics are stored per run. Open run history, inspect metrics and checkpoints, pick a model for promotion to a capital deployment, and keep a clear audit trail of what was trained and when.