---

name: keras-rl description: "Use legacy keras-rl reinforcement-learning agents, memories, policies, processors, callbacks, and safe smoke checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

keras-rl

Use this repo skill when a task involves the legacy keras-rl package: Keras 2.x reinforcement-learning agents, Gym-style environments, DQN/Double DQN/Dueling DQN, SARSA, CEM, DDPG, NAF/CDQN, replay memories, exploration policies, processors, callbacks, logging, or lightweight compatibility checks.

First checks

• Treat keras-rl as legacy standalone Keras 2.x code. Do not assume modern Keras 3 or pure tf.keras compatibility.
• Prefer a Keras-2-compatible backend stack for any real execution. A Theano CPU backend is often the safest compile-only route for this package family; TensorFlow-era imports can succeed while agent construction still fails on legacy symbolic-tensor assumptions.
• Install the package itself with one of:

pip install keras-rl
pip install "keras-rl[gym]"  # when Gym examples or environments are needed

Additional workflow dependencies are separate: h5py for weight files, matplotlib for log plotting, wandb for WandbLogger, Atari extras/ROMs for Atari tasks, and MuJoCo system/license dependencies for MuJoCo tasks.

Minimal import check:

import rl
from rl.agents import DQNAgent, DDPGAgent, CEMAgent, SARSAAgent, NAFAgent
from rl.memory import SequentialMemory, EpisodeParameterMemory
from rl.policy import EpsGreedyQPolicy, BoltzmannQPolicy

For environment and backend diagnostics, use the checker linked from core-extension-and-logging before spending time on training.

Route by task

Discrete-action agents

Use discrete-control when the task names or implies:

• DQN, Double DQN, Dueling DQN, SARSA, CEM, CartPole, Atari-style DQN, or a discrete Gym action space.
• SequentialMemory, EpisodeParameterMemory, EpsGreedyQPolicy, BoltzmannQPolicy, LinearAnnealedPolicy, replay warmup, target networks, or dueling aggregation.
• A need to compile a safe discrete-agent smoke script without long training or rendering.

Continuous-action agents

Use continuous-control when the task names or implies:

• DDPG, NAF/CDQN, Pendulum, MuJoCo, continuous action spaces, actor/critic wiring, critic_action_input, V_model, mu_model, L_model, or covariance modes.
• OrnsteinUhlenbeckProcess, GaussianWhiteNoiseProcess, continuous replay memory, action-noise sizing, or a compile-only continuous-agent smoke check.

Core lifecycle, processors, callbacks, and logging

Use core-extension-and-logging when the task involves:

• Agent.fit, Agent.test, compile-before-fit errors, callback wiring, verbose, visualize, nb_max_episode_steps, action_repetition, or start_step_policy.
• Custom Processor or Env implementations, MultiInputProcessor, WhiteningNormalizerProcessor, Gym API adapters, or state-batch shape transformations.
• FileLogger, ModelIntervalCheckpoint, WandbLogger, log JSON plotting, environment compatibility checks, clone_model, clone_optimizer, huber_loss, or WhiteningNormalizer.

Cross-cutting references

• Read Repository Provenance before deciding whether this skill is current for a different checkout or after package code changes.
• Read Install and compatibility when imports, Keras backend setup, Gym optional dependencies, WandB, matplotlib, weight files, Atari, or MuJoCo requirements are uncertain.
• Read Troubleshooting for package-wide failure triage before drilling into a sub-skill-specific troubleshooting page.

Safe operating stance

• Prefer compile/build smoke checks before fit or test. Full training examples can be long-running, display-dependent, or require external simulators.
• Keep Atari and MuJoCo tasks explicitly optional unless the user has already installed the required extras, ROMs/data, system packages, and licenses.
• Do not use original repository examples as runtime dependencies. Use the bundled smoke helpers and references in this skill tree instead.
• For any workflow that spans agent construction and logging/lifecycle, start in the agent-family sub-skill, then cross-check Agent.fit/callbacks in core-extension-and-logging.