Feature: Add Reinforcement Learning Support

[orlandohohmeier]

Summary

Implement Reinforcement Learning (RL) support based on the approach outlined in RFC #69.
This feature builds on additions to the scheduler and worker configuration, particularly:

• Task #80 — Driver Configuration Support, providing the flexibility to define custom drivers for RL components.
• Task #76 — Externalized Scheduler Configuration, enabling dynamic scheduling setups that can accommodate RL workflows.

The goal is to provide the necessary components and integration logic for RL training loops within the Hypha framework, ensuring compatibility with the training requirements of the SPRIN-D Composite Learning Challenge as outlined in Issue #53.

Key goals:

• Implement RL-specific drivers and configurations (e.g., policy learner, environment simulator, experience aggregator).
• Define a reusable scheduling configuration for RL tasks using the external scheduler configuration mechanism.
• Ensure drivers interoperate correctly with the Hypha scheduler for data and experience flow.
• Enable RL training workflows that align with SPRIND Challenge requirements.

Background

The implementation of Reinforcement Learning support is a critical next step for enabling model training workflows. The RFC (#69) describes a modular RL architecture where workers run different RL roles (e.g., policy learner, environment simulator, aggregator) and coordinate through Hypha’s decentralized scheduler.

Configurable drivers (#80) and externalized scheduler setups (#76) will provide the foundation to easily define and deploy RL-specific drivers.

This feature will integrate these components to enable RL task orchestration, ensuring full end-to-end support for RL-based model training workflows.