hpo-and-mutation

# AgileRL HPO And Mutation Use this sub-skill when the task is about AgileRL evolutionary hyperparameter optimization (HPO), `TournamentSelection`, `Mutations`, `HyperparameterConfig`, `RLParameter`, mutation probabilities, architecture mutation behavior, or diagnosing why a population did or did not mutate. ## Read First - `references/hpo-workflows.md` for the standard tournament + mutation loop. - `references/api-reference.md` for key HPO classes and parameters. - `references/troubleshooting.md` for invalid mutable attributes, probability, and architecture mutation failures. - `scripts/inspect_hpo_setup.py --help` for a safe config-only HPO probe. ## Boundaries - Training-loop placement belongs in `../training-workflows/SKILL.md`, `../multi-agent-and-wrappers/SKILL.md`, `../offline-bandits-data/SKILL.md`, or `../llm-fine-tuning/SKILL.md` depending on the workflow. - Architecture object construction belongs in `../evolvable-modules/SKILL.md`. - This sub-skill owns which hyperparameters are mutable, how mutation probabilities are selected, how tournament selection preserves elites, and why mutation may be a no-op. ## Standard HPO Flow 1. Create a population of AgileRL agents. 2. Evaluate population fitness. 3. Select elites and the next generation with `TournamentSelection`. 4. Mutate offspring with `Mutations`. 5. Continue training with the evolved population. ```python from agilerl.hpo.tournament import TournamentSelection from agilerl.hpo.mutation import Mutations from agilerl.algorithms.core.registry import HyperparameterConfig, RLParameter hp_config = HyperparameterConfig( lr=RLParameter(min=1e-4, max=1e-2), batch_size=RLParameter(min=32, max=256), learn_step=RLParameter(min=1, max=10, grow_factor=1.5, shrink_factor=0.75), ) tournament = TournamentSelection(tournament_size=2, elitism=True, population_size=6, eval_loop=1) mutations = Mutations( no_mutation=0.4, architecture=0.2, new_layer_prob=0.2, parameters=0.2, activation=0.0, rl_hp=0.2, mutation_sd=0.1, rand_seed=1, device="cpu", ) ``` ## Decision Points - Set `activation=0` when activation mutations are unsupported or not desired. - Use `HyperparameterConfig` fields that exactly match algorithm attributes. - Keep `population_size`, `POP_SIZE`, and tournament population size aligned. - Use deterministic seeds when comparing HPO behavior. - For LLM algorithms, do not assume architecture mutations are supported.