---

name: continuous-control description: "Build, configure, troubleshoot, and lightly verify continuous-action keras-rl DDPG and NAF workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

continuous-control

Use this sub-skill when a task involves continuous-action keras-rl agents: DDPG, NAF/CDQN, actor-critic wiring, critic_action_input, value/mu/L model construction, OrnsteinUhlenbeckProcess, GaussianWhiteNoiseProcess, Pendulum-style control, or MuJoCo caveats.

Route here for

• DDPG actor/critic model construction and DDPGAgent.compile([actor_optimizer, critic_optimizer]) setup.
• NAF/CDQN V_model, mu_model, L_model, covariance_mode, and continuous-action assumptions.
• Replay memory and action-noise sizing as they affect continuous agents.
• Safe compile-only checks for installed keras-rl continuous-agent wiring.
• Pendulum-style Gym tasks and reference-only MuJoCo adaptations.

Route elsewhere

• DQN, Double/Dueling DQN, SARSA, or CEM discrete-action recipes: use discrete-control.
• The shared Agent.fit/Agent.test lifecycle, custom Processor base classes, callbacks, logging, or visualization helpers: use core-extension-and-logging.

Use these bundled resources

1. API reference — constructor signatures, model shape contracts, random processes, and memory notes.
2. Workflows — DDPG and NAF build recipes, compile smoke usage, Pendulum guidance, and MuJoCo caveats.
3. Troubleshooting — shape, optimizer, covariance, backend, Gym, MuJoCo, and weight-file failures.
4. Continuous-agent smoke script — compile/build-only helper for DDPG and NAF from an installed keras-rl package.

Operating notes

• Treat keras-rl as legacy standalone Keras 2.x code. Prefer a legacy Keras backend and check package compatibility before assuming modern TensorFlow/Keras behavior.
• Keep continuous-action examples bounded: compile/build smokes are safe; long training, display rendering, MuJoCo setup, and weight persistence are opt-in user tasks.
• For any continuous Gym environment, first extract nb_actions from a one-dimensional continuous action space and keep every actor/mu/random-process output shaped (nb_actions,) per action sample.