Imitation learning

decisionrl.imitation learns policies from demonstrations rather than rewards.

Method	Class	Idea
Behavioral Cloning	BC	supervised action prediction from a demo dataset
DAgger	DAgger	roll out the policy, relabel visited states with an expert, aggregate, retrain
GAIL	GAIL	a discriminator separates expert vs policy transitions; the policy (PPO) is trained to fool it

from decisionrl.imitation import BC, DAgger, GAIL, collect_expert_dataset
from decisionrl.envs import CartPole

demos = collect_expert_dataset(CartPole(), expert_policy, n_transitions=4000, seed=0)

# Behavioral cloning — pure supervised imitation
bc = BC(CartPole(), seed=0)
bc.train(demos, n_iters=1500)

# DAgger — needs a queryable expert to fix compounding error
dagger = DAgger(CartPole(), seed=0)
dagger.learn_dagger(CartPole(), expert_policy, iterations=4)

# GAIL — adversarial imitation, no environment reward at all
gail = GAIL(CartPole(), demos, seed=0)
gail.learn(iterations=10)

On CartPole with a heuristic expert, BC and DAgger reach the maximum return (500); GAIL matches the expert from demonstrations alone, never seeing a reward. Complements the offline-RL agents (TD3BC, IQL, CQL, DecisionTransformer) and the preference-based methods (decisionrl.rlhf).