testing_guide.md

OptimusEnv Testing and Verification Guide

This guide provides step-by-step instructions for testing, verifying, and interacting with the OptimusEnv reinforcement learning environment.

📱 Live Environment

The official OpenEnv-compliant endpoint is hosted at: https://manasdutta04-optimusenv.hf.space/

---

1. Quick Health Check (CURL)

To verify the environment is online and responsive:

curl https://manasdutta04-optimusenv.hf.space/health

Expected Output: {"status":"ok", "env":"OptimusEnv", "version":"1.0.0"}

---

2. Inspecting the Task Suite

List all available RL tasks, including their descriptions and required action schemas:

curl https://manasdutta04-optimusenv.hf.space/tasks

---

3. Running an Episode (Step-by-Step)

You can manually interact with the environment to verify the transition logic:

Step 1: Start/Reset an Episode

Choose a task (e.g., task_1):

curl -X POST https://manasdutta04-optimusenv.hf.space/reset \
  -H "Content-Type: application/json" \
  -d '{"task":"task_1"}'

Step 2: Perform an Action

Run one training epoch with specific hyperparameters:

curl -X POST https://manasdutta04-optimusenv.hf.space/step \
  -H "Content-Type: application/json" \
  -d '{
    "learning_rate": 0.001,
    "batch_size": 64,
    "optimizer": "adamw",
    "num_layers": 2,
    "hidden_dim": 128,
    "use_amp": false,
    "lr_schedule": "cosine",
    "weight_decay": 0.0001
  }'

---

4. Comprehensive Testing Scripts

For full execution verification, use the provided Python scripts in the repository.

A. Run the Heuristic Baseline

This runs a pre-defined policy across all three tasks and reports a final score:

# Point the script to the live URL
python baseline/run_baseline.py --host https://manasdutta04-optimusenv.hf.space

B. Run the Submission Inference

Simulate an agent run (this uses the same interface as the hackathon validator):

export API_BASE_URL=https://manasdutta04-optimusenv.hf.space
python inference.py

---

💡 Troubleshooting and Tips

Dataset Pre-caching: The first time a task is run, the environment may take a few seconds to download the relevant PyTorch datasets (MNIST, CIFAR-10, etc.). Subsequent steps/resets will be much faster.

Port Consistency: If running locally, the environment serves on port 8000. On Hugging Face Spaces, use the root URL provided.

Crash Penalty: Running an action with extremely high learning rates or too many layers (causing OOM) will result in a reward of -1.0 and terminate the episode immediately.

📊 Interpreting Results

• /grader: This endpoint returns the final normalized score [0.0 - 1.0] for the last completed episode.
• /baseline: This runs a server-side heuristic test to provide a reference score for all tasks.

---

For more technical details on the reward function and observation space, refer to the README.md.