We're looking forward to seeing you all at @RL_Conference tomorrow! 🚀.📍CCIS 1-440 🕘 starting 9am.🔗get the full schedule here:

We will conclude with a discussion of alternative assumptions which might be more friendly to the development of agents capable of the open-ended, iterative, and creative intelligent behaviours necessary for pushing the frontier of human knowledge.

such as the distinction between the agent and environment, the stationarity of the environment transition dynamics, and the external source of reward, influence the types of behaviour we can expect to observe in our agents.

is it reasonable to explain humanity’s greatest scientific discoveries or works of art with the same framework that describes the neural process of training a dog to sit? This talk will explore how fundamental assumptions that underlie the standard RL problem formulation,.

However, we have yet to see RL agents propose novel paradigms of mathematics or write great works of literature, begging the question:.

Reinforcement learning has proven to be an outrageously effective framework in settings where a) reward is well-defined and b) it is possible to quickly collect a diverse set of experiential data containing a gradient of reward signals.

Completing our fantastic lineup of invited talks is Dr. @clarelyle (Deepmind) who will speak on Beyond Optimality: Designing Open-Ended RL Agents. Abstract in 🧵

I will discuss some of the reasons why RL might want to take minds seriously, despite some of the challenges involved in studying them.

Reinforcement learning studies systems that generate reward-maximizing behavior. Where does the idea of a mind fit into this picture, if it’s even needed? Should we be behaviorists, cognitivists, or something in between?.

Professor @mark_ho_ (NYU) will give an invited talk on Why Having a Mind Might be Useful 🧠 Abstract in 🧵

Finally, I will briefly discuss how 4) a reinforcement-learning agent able to do so can access the rest of AI.

and that therefore 3) a critical component for achieving general intelligence is the ability to autonomously construct task-specific frames, which can be achieved by learning mutually-compatible observation and action abstractions.

I will argue that: 1) the right outer frame for a generally-intelligent agent is a decision process, 2) that that decision process must necessarily be drastically overpowered, compared to the natural framing of any individual task it may wish to solve,.

Professor George Konidaris (Brown University) will be giving an invited talk titled: Agents Must Learn Their Own Frames. Abstract in 🧵

ending on the hopeful note that maybe, just maybe, some of the right questions are starting to come into view.

And yet, making sustained, robust progress toward this goal has proved surprisingly, infuriatingly difficult. We'll take a high-level tour of some garden paths and blind alleys that have caused useful shifts in my own frame,.

Model-based reinforcement learning (MBRL), i.e. the problem of learning a predictive model and using it to plan, seems like a straightforward and intuitively sensible algorithmic proposition. Doing it well could transform the way we approach many core problems in RL.

We have a really exciting lineup of invited speakers this year 🔥 Kicking us off we have Prof. Erin Talvitie (Harvey Mudd College), whose talk is titled: 20 Years of Asking the Wrong Questions in Model-based reinforcement learning. Abstract in 🧵

The Most Thought-Provoking paper award goes to Thinking is Another Form of Control 🏆 Congratulations to  @JosiahHanna and Nicholas E. Corrado! 🎉 Check out their paper here:

🔎Among many fantastic submissions, some especially stood out: we will hear contributed talks on Thinking is Another Form of Control, Analogy Making as Amortised Model Construction, and Agent-centric Learning: from External Reward Maximization to Internal Knowledge Curation✨.