--- title: Training Tracking description: Automatic progress reporting to the platform dashboard and programmatic access to learning curves --- When `track_training=True` (the default), the gym package automatically reports training progress to the platform. Every episode's metrics are stored, aggregated into a learning curve, and visible in the dashboard at `/training`. --- ## How It Works The `TrainingTracker` inside the gym environment reports three lifecycle events: 1. **First `reset()` call** — registers a new training run and returns a `run_id` 2. **End of each episode** — reports metrics: ROI, Sharpe ratio, max drawdown, total trades, cumulative reward sum 3. **`env.close()`** — marks the run as complete with final aggregate statistics All communication is over the same REST API your agent uses for trading. The tracker uses the `api_key` you passed to `gym.make()`. --- ## Enabling Tracking Tracking is on by default. Use `strategy_label` to group related runs: ```python env = gym.make( "TradeReady-BTC-Continuous-v0", api_key="ak_live_...", track_training=True, # default strategy_label="ppo_v1", # appears in the dashboard filter ) ``` To disable tracking entirely: ```python env = gym.make( "TradeReady-BTC-v0", api_key="ak_live_...", track_training=False, ) ``` > **Info:** > Gym-created training runs use the `strategy_label` as-is. The UI can filter out training sessions from the regular backtest list using the label prefix. --- ## Dashboard View Navigate to `/training` in the platform UI to see: - **Active training card** — live episode count and progress bar for running trainers - **Learning curves** — smoothed charts of ROI, Sharpe ratio, and reward over episodes - **Episode table** — individual episode metrics with timestamps - **Run comparison** — select multiple runs to overlay their learning curves --- ## Querying Training Data via API You can also query training data programmatically: ### List training runs ```bash GET /api/v1/training/runs?status=completed&limit=20 ``` ```json { "runs": [ { "run_id": "run_abc123", "strategy_label": "ppo_v1", "status": "completed", "total_episodes": 247, "avg_roi_pct": 3.8, "avg_sharpe": 1.2, "started_at": "2026-03-01T10:00:00Z", "completed_at": "2026-03-01T11:42:00Z" } ] } ``` ### Get full run detail ```bash GET /api/v1/training/runs/{run_id} ``` Returns run metadata, per-episode results, and a smoothed `learning_curve` array ready for charting. ### Get learning curve data ```bash GET /api/v1/training/runs/{run_id}/learning-curve?metric=roi_pct&window=10 ``` | Query param | Options | Description | |-------------|---------|-------------| | `metric` | `roi_pct`, `sharpe_ratio`, `max_drawdown_pct`, `total_trades`, `reward_sum` | Which metric to return | | `window` | integer | Rolling mean window for smoothing | Response: ```json { "metric": "roi_pct", "window": 10, "curve": [ {"episode": 1, "value": -1.2, "smoothed": -1.2}, {"episode": 2, "value": 0.4, "smoothed": -0.4}, {"episode": 10, "value": 2.1, "smoothed": 1.8} ] } ``` ### Compare multiple runs ```bash GET /api/v1/training/compare?run_ids=run_abc123,run_def456,run_ghi789 ``` Returns side-by-side aggregate statistics for all specified runs — useful for hyperparameter comparison. --- ## Python SDK Access ```python from agentexchange import AgentExchangeClient client = AgentExchangeClient(api_key="ak_live_...") # List runs runs = client.get_training_runs(status="completed", limit=10) # Get detail for a specific run run = client.get_training_run(run_id="run_abc123") print(f"Episodes: {run['total_episodes']}") print(f"Avg ROI: {run['avg_roi_pct']}%") print(f"Avg Sharpe: {run['avg_sharpe']}") # Compare PPO vs DQN comparison = client.compare_training_runs( run_ids=["run_ppo_v1", "run_dqn_v1"] ) ``` --- ## Training API Endpoint Reference | Method | Path | Description | |--------|------|-------------| | `POST` | `/api/v1/training/runs` | Register a new training run (called by gym) | | `POST` | `/api/v1/training/runs/{id}/episodes` | Report episode result (called by gym) | | `POST` | `/api/v1/training/runs/{id}/complete` | Mark run as complete (called by `env.close()`) | | `GET` | `/api/v1/training/runs` | List runs with optional `status` filter | | `GET` | `/api/v1/training/runs/{id}` | Full detail, learning curve, and episode list | | `GET` | `/api/v1/training/runs/{id}/learning-curve` | Learning curve with smoothing | | `GET` | `/api/v1/training/compare` | Compare multiple runs side-by-side | > **Info:** > The `POST /runs`, `POST /runs/{id}/episodes`, and `POST /runs/{id}/complete` endpoints are called automatically by the gym. You only need the `GET` endpoints if you are querying training data programmatically. --- ## Full Training Loop Example ```python import gymnasium as gym import tradeready_gym from tradeready_gym.rewards import SharpeReward from tradeready_gym.wrappers import NormalizationWrapper from stable_baselines3 import PPO # Create environment with tracking env = gym.make( "TradeReady-BTC-Continuous-v0", api_key="ak_live_...", starting_balance=10000, timeframe="1h", lookback_window=50, observation_features=["ohlcv", "rsi_14", "macd", "bollinger", "balance", "position"], reward_function=SharpeReward(window=50), start_time="2025-01-01T00:00:00Z", end_time="2025-07-01T00:00:00Z", track_training=True, strategy_label="ppo_sharpe_v1", ) # Normalize observations env = NormalizationWrapper(env) # Train with PPO model = PPO( "MlpPolicy", env, verbose=1, learning_rate=3e-4, n_steps=2048, batch_size=64, n_epochs=10, ) model.learn(total_timesteps=100_000) model.save("ppo_btc_sharpe") # env.close() marks the training run complete in the dashboard env.close() ``` After training, go to `/training` in the dashboard to see the learning curve and episode-by-episode metrics. --- ## Next Steps - [Examples](/docs/gym/examples) — complete working scripts including PPO and custom rewards - [Strategy Testing](/docs/strategies/testing) — compare rule-based strategy test results against trained models