evolvable-modules

# AgileRL Evolvable Modules Use this sub-skill when a task asks for AgileRL `EvolvableModule`, `EvolvableNetwork`, `net_config`, MLP/CNN/LSTM/MultiInput/SimBa configuration, Q/value/actor networks, custom architectures, or wrapping ordinary PyTorch modules for AgileRL mutation compatibility. ## Read First - `references/modules-and-networks.md` for conceptual relationships and build patterns. - `references/api-reference.md` for key classes and config objects. - `references/configuration.md` for `encoder_config`, `head_config`, image/recurrent/multi-input settings. - `references/troubleshooting.md` for shape, mutation, and custom module errors. - `scripts/inspect_evolvable_builders.py --help` for safe tiny builder checks. ## Boundaries - Use `../training-workflows/SKILL.md` for full Gymnasium training loops. - Use `../hpo-and-mutation/SKILL.md` for mutation probabilities and tournament selection. - Use `../multi-agent-and-wrappers/SKILL.md` for grouped multi-agent `net_config` and PettingZoo agent IDs. - This sub-skill owns architecture selection, config validation, custom wrappers, and network construction. ## Common Routes | Task | Guidance | | --- | --- | | Vector observation to discrete action values | Use `QNetwork` with MLP encoder/head config. | | Image observation policy/value network | Use CNN encoder config and verify channels/order. | | Dict/Tuple observations | Use MultiInput config with MLP/CNN/LSTM sub-configs as needed. | | Recurrent/partially observable workflow | Use LSTM config and read training recurrent notes. | | Custom PyTorch module | Inherit `EvolvableModule` when architecture should mutate, or use `DummyEvolvable` when only RL hyperparameters/weights should mutate. | | SimBa/ResNet style architecture | Use the corresponding module/network configs and validate dimensions before training. | ## Safe Validation ```bash python scripts/inspect_evolvable_builders.py --mode mlp python scripts/inspect_evolvable_builders.py --mode dict ``` The helper constructs tiny Gymnasium spaces and config objects. It does not train.