Rl

The authors propose PQN, a simplified version of DQN that leverages LayerNorm and L2 regularization to get rid of the target networks while maintaining a high level of training stability. The authors also prove a convergence property on PQN, which wasn’t available for the current de-facto algorithm PPO.

Review of a MARL algorithm that enables learning a set of different policies while sharing gradients from a policy to all others. Thus enabling the agents to learn from it’s own on-policy data and also from off-policy data coming from other agents in that environment.

Review of the original technique which is widely used today for LLM alignment. It’s the foundation to many other alignment techinques proposed in the recent litterature.