The authors derive a relationship between empowerment and mutual-information skill learning, a family of methods that trains distinguishable behaviors without task rewards.
Agents keep more options open by seeking states rich in skills
The authors turn scalable skill-discovery algorithms into policies that seek controllable states, producing resource gathering, survival and construction behaviors.
Top University
Catherine Ji · Vivek Myers · Joshua Lin · Alexandra Wingate · Sergey Levine · Benjamin Eysenbach
Princeton University · UC Berkeley
Research Digest··3 min read
Ji and colleagues connect empowerment, an agent’s capacity to control future outcomes, with unsupervised skill learning.
Why this paper
From Princeton University and UC Berkeley
In one line
Empowerment maximization can be scaled using skill learning methods, yielding emergent behaviors like building houses in Crafter.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§