Algorithms

AgileRL already includes state-of-the-art evolvable on-policy, off-policy, offline and multi-agent reinforcement learning algorithms with distributed training. We are constantly adding more algorithms, with a view to add hierarchical algorithms soon.

Core algorithm tools:

• EvolvableAlgorithm
• Algorithms Mutations Registry
• OptimizerWrapper

Observation Spaces

AgileRL currently supports processing the following observation spaces:

Box (Continuous)	Discrete	Dict	Tuple
✅	✅	✅	✅

Single-Agent Algorithms

Action Spaces

Algorithm	Discrete	Box	MultiDiscrete	MultiBinary
CQL	✅	❌	❌	❌
DDPG	❌	✅	❌	❌
DQN	✅	❌	❌	❌
DQN Rainbow	✅	❌	❌	❌
ILQL	✅	❌	❌	❌
PPO	✅	✅	✅	✅
TD3	❌	✅	❌	❌

Multi-Agent Algorithms

Action Spaces

Algorithm	Discrete	Box	MultiDiscrete	MultiBinary
IPPO	✅	✅	✅	✅
MADDPG	✅	✅	❌	❌
MATD3	✅	✅	❌	❌

Bandit Algorithms

• Neural Contextual Bandits with UCB-based Exploration (NeuralUCB)
• Neural Contextual Bandits with Thompson Sampling (NeuralTS)

LLM Algorithms

Algorithm	Training Mode	Notes
GRPO	On-policy RL	Group-relative policy optimization for LLMs.
CISPO	On-policy RL	GRPO specialization with fixed CISPO loss objective.
GSPO	On-policy RL	GRPO specialization with fixed GSPO sequence-level objective.
DPO	Preference optimization	Offline preference learning without reward-model RL loop.
SFT	Supervised fine-tuning	Cross-entropy on response tokens; typical warm-up before DPO/RL.
LLM PPO	On-policy RL	PPO-style actor-critic objective for token-level trajectories.
LLM REINFORCE	On-policy RL	REINFORCE-style policy-gradient objective for LLM finetuning.

• Group Relative Policy Optimization (GRPO)
• Clipped Importance Sampling Policy Optimization (CISPO)
• Group Sequence Policy Optimization (GSPO)
• Direct Preference Optimization (DPO)
• Supervised Fine-Tuning (SFT)
• LLM Proximal Policy Optimization (LLM PPO)
• LLM REINFORCE
• Supervised Fine-Tuning (SFT)