Core Concepts
Complexity is an illusion created by abstraction, as in the absence of abstraction layers, all behaviors have equal complexity. However, in the context of spatially and temporally extended abstraction layers, efficiency demands that weak constraints take simple forms, leading to a correlation between simplicity and sample efficiency, which is not a causal relationship but rather a result of confounding.
Abstract
The content explores the concept of complexity and its relationship to simplicity and sample efficiency in machine learning and artificial intelligence.
Key points:
In the absence of abstraction layers, the complexity of all behaviors is equal, suggesting that complexity is a subjective "illusion" rather than an objective property.
When considering finite vocabularies in spatially and temporally extended environments, the author argues that policy weakness can confound sample efficiency with policy simplicity.
This is because goal-directed abstraction tends to favor weak constraints that take simple forms, as a larger vocabulary exponentially increases the space of outputs and policies, which may conflict with finite time and space constraints.
The author suggests that the correlation between simplicity and sample efficiency is not a causal relationship, but rather a result of this confounding effect.
The content builds upon previous work that showed maximizing policy "weakness" is necessary and sufficient to maximize sample efficiency, and that experiments demonstrated weak policies outperforming simple ones by 110-500%.