What this project is

A Double DQN implementation and evaluation pipeline for analysing value overestimation and training stability.

Why it matters

Many value-based methods can look stable while quietly accumulating biased estimates. Double DQN offers a direct way to reduce this bias.

Architecture overview

  1. Replay buffer and target network pipeline.
  2. Double DQN target calculation.
  3. Training/evaluation split with fixed seeds.
  4. Monitoring of value estimates and policy performance.

Current status

  • Baseline training loop and replay handling implemented.
  • Initial comparisons against vanilla DQN complete.
  • Experiment tracking templates established.

Selected results

  • Reduced overestimation relative to vanilla baseline in early experiments.

Roadmap

  • Expand ablations over replay and target-update frequency.
  • Improve run-to-run variance analysis.