Command-line interface

Installing the package provides a decisionrl console command (and python -m decisionrl).

decisionrl list                                          # algorithms & environments
decisionrl train ppo CartPole --steps 50000 --save ppo.pt
decisionrl train dqn CartPole --set learning_rate=5e-4 --set buffer_size=100000
decisionrl eval ppo --env CartPole --load ppo.pt --episodes 20
decisionrl train ppo gym:LunarLander-v2 --steps 200000   # any Gymnasium env

Tuned defaults

train applies tuned default hyperparameters per (algorithm, environment) automatically. Override any of them with repeated --set KEY=VALUE, or disable them with --no-tuned.

Programmatic registry

The same string-based construction is available in code:

from decisionrl import make_env, make_agent, list_algorithms

print(list_algorithms())
agent = make_agent("ppo", make_env("CartPole"), seed=0)
agent.learn(50_000)