Summary
Experiment with more advanced network backbones (e.g., Residual blocks, Squeeze-and-Excitation) for the policy/value heads.
Motivation
Stronger architectures may improve the model’s learning capacity and game strength.
Acceptance criteria / Definition of Done
- Create a modular network construction (e.g.,
models/architectures.py) supporting at least:
- Baseline small CNN (existing).
- One ResNet-style variant (configurable via
config.yaml).
- Provide scripts / instructions to train or load the alternative architecture.
- Report a short comparison (loss/validation metrics or small head-to-head) in
docs/ showing any measurable differences or at least proving the architecture can train.
Implementation notes
- Keep versions small and parameter-controlled so experiments remain tractable.
- Ensure architecture choice is exposed in the config file (see config issue).
Complexity
Advanced — good for contributors with deep learning experience.
Summary
Experiment with more advanced network backbones (e.g., Residual blocks, Squeeze-and-Excitation) for the policy/value heads.
Motivation
Stronger architectures may improve the model’s learning capacity and game strength.
Acceptance criteria / Definition of Done
models/architectures.py) supporting at least:config.yaml).docs/showing any measurable differences or at least proving the architecture can train.Implementation notes
Complexity
Advanced — good for contributors with deep learning experience.