[Feature] Implicit Q Learning

[cvnad1]

Hi,

I would like to add the Implicit Q-Learning algorithm to the Acme framework.

https://arxiv.org/abs/2110.06169

[AryanVBW]

Thanks for the feature request! Implicit Q-Learning (IQL) is a great offline RL algorithm.

If you'd like to contribute an implementation:

1. Check the CONTRIBUTING.md for guidelines
2. Look at existing offline agents like CQL/TD3-BC in acme/agents/jax/ as reference implementations
3. IQL key components needed:
   • Expectile regression for value function
   • AWR-style policy extraction
   • Standard Q-function training

Reference implementations you could adapt:

• Original IQL: https://github.com/ikostrikov/implicit_q_learning
• JAX IQL: https://github.com/rail-berkeley/rlkit

Note: There was a prior discussion in #204 about contributing TD3-BC and IQL implementations. The codebase structure for offline agents is already in place.

If you open a PR, the maintainers can provide feedback on integration with Acme's builder pattern.