actor_learner_protocolValidate IMPALA actor learner unrolls with behavior and learner policy metadata for V-trace recovery.
Independent GitHub snapshot · Automated classification · Original documentation governs
Use this skill when constructing or checking IMPALA-style trajectories before V-trace computation. It should not compute losses or update parameters.
Inputs: rewards, discounts, actions, values including bootstrap, learner policy distributions, and behavior policy distributions. Outputs: validated JSON with action probabilities and policy-lag diagnostics.
Workflow: load an unroll, check dimensions and probability distributions, extract probabilities for each action, and emit a compact validated record. Preserve both learner and behavior policy metadata because treating the sample as on-policy violates the paper mechanism.
Validation: run python tests/test_protocol.py or validate the skill tree with tests. Limitations: this skill uses small JSON fixtures and does not launch distributed actors.