Together, these contributions demonstrate how extended action representations and advanced policy models can advance the efficiency and versatility of RL.

Finally, we present diffusion policies as a more expressive policy class for maximum entropy RL, and highlight their advantageous properties for stability, flexibility, and scalability in complex domains.

Building on this foundation, we introduce a novel algorithm for skill discovery with MPs that leverages maximum entropy RL and mixture-of-expert models to autonomously acquire diverse, reusable skills.

However, standard MP-based approaches result in open-loop policies; to address this, we extend them with online replanning of MP trajectories and off-policy learning strategies that exploit single-time step information.

This parametrization allows black-box RL algorithms to adapt MP parameters to diverse contexts and initial states, providing a pathway toward versatile skill acquisition.

they encode trajectories with a concise set of parameters, naturally yielding smooth behaviors and enabling exploration in parameter space rather than in raw action space.

Abstract: Reinforcement learning (RL) with primitive actions often leads to inefficient exploration and brittle behaviors. Extended action representations, such as motion primitives (MPs), offer a more structured approach:

I will suggest a framework for answering these questions
through the medium of potential-based shaping - in which 'liking'
provides immediate, but preliminary and ultimately cancellable,
information about the true, long-run worth of outcomes.

How could it be that
we, or an agent, could `want' something that it does not `like', or
`like' something that it would not be willing to exert any effort to
acquire?

I will talk about an
example of the complexity that has important psychological and neural
resonance - namely the distinct concepts of 'liking' and 'wanting'. The
former characterizes an immediate hedonic experience; and the latter the
motivational force associated with that experience.

Abstract: As reinforcement learners, humans and other animals are excellent at
improving their otherwise miserable lot in life. This is often described
in terms of optimizing utility. However, understanding utility in a
non-circular manner is surprisingly difficult.

This talk focuses on recent research advances from the Game Intelligence team at Microsoft Research, towards scalable machine learning architectures that effectively model human gameplay, and our vision of how these innovations could empower creatives in the future.

Progress towards this goal has exciting potential for applications in video games, from new tools that empower game developers to realize new creative visions, to enabling new kinds of immersive player experiences.

Abstract: Modeling complex environments and realistic human behaviors within them is a key goal of artificial intelligence research.