To move fast, we are splitting the work into two tracks. We use the Gymnasium interface as our "handshake" to ensure our code connects perfectly.
- [cite_start]Topology: Build the IBM Falcon 27q graph and distance matrix[cite: 10].
- [cite_start]Circuit Logic: Use Qiskit to load QASMBench and find the "front layer" of gates[cite: 10].
- Step Function: Define how the mapping changes after a SWAP.
- [cite_start]Done Criteria: Signal when all gates are executed[cite: 10].
- [cite_start]The Model: Set up the DQN or Maskable PPO architecture[cite: 12].
- [cite_start]Action Masking: Prevent the agent from choosing invalid SWAPs[cite: 10].
- [cite_start]Reward Logic: Implement the scoring system to drive learning[cite: 10].
- [cite_start]Hyperparameters: Tune the lookahead window (
$K$ ) and learning rates[cite: 10].
- We agree on the State Shape (how the mapping and gates are represented).
- We use Stable Baselines3 to keep the training code standard.