On October 20th and 21st 2025, the School for Information and Knowledge Systems (SIKS) will be organising a new two-day course on  Reinforcement Learning for Adaptive Hybrid Intelligence. If you are interested, please do reserve these dates. Registration will open soon via: https://siks.nl/activities/activities/siks-course-rl-for-ahi-2025/ (also see that website for more information).

 

Recent years have seen notable breakthroughs in reinforcement learning in robotics, games such as Atari and Go, and in ChatGPT. While the focus on autonomous and active learning makes reinforcement learning a powerful tool, deployment of a reinforcement learning agent as an assistant or collaborator (that is, as a hybrid intelligence) raises specific challenges.

In this two-day course we will take a look at the foundations of reinforcement learning and various extensions that are important for hybrid intelligence. These include:

  • “Safe reinforcement learning”: safety, ethical, or legal constraints can shape the reward function and eventually help the artificial agent in exploration vs. exploitation trade-off
  • “Causal reinforcement learning” will take on designing RL agents that can interact with the environment in both explaining the data generation process in the environment, and making better decisions provided that they are equipped with causal knowledge
  • “Learning from feedback” will focus on obtaining behavior or reward functions from data or interactions where reward functions are  hard to specify manually
  • “RL in the context of LLMs” will focus on preference optimization,  RLHF  and the GRPO algorithm in order to adjust for human preferences, with a hands-on  demo.
  • “Multi-agent RL (MARL)” will focus on the setting where multiple RL agents interact in a shared environment, requiring communication and coordination.
  • “Safe MARL”:  While every agent has an individual safety guarantee,  the overall multiagent system can still be risky due to coordination/cooperation issues.  This component will address such challenges in addition to maintaining  safe exploration, using formal constraints in multiagent environments.
  • “Multi-objective RL” will focus on scenarios where the agent should take into account multiple, potentially conflicting, criteria and goals.

We hope to see many of you there!

Erman Acar, Shihan Wang & Herke van Hoof