Successor features — what and why?

hard

Answer

$Barreto et al. 2017.$
$Decompose reward as r = φ(s, a, s') \cdot w.$
${\psi}(s)\; = \;E[{\Sigma}\;{\gamma}^{k}\;{\varphi}(s_{k}, \;a_{k}, \;sk)]$
$New task (new w) can reuse learned ψ — instant transfer without relearning value function.$
$Foundation of generalized policy improvement (GPI) across tasks; applications in continual learning + universal RL.$

Check yourself — multiple choice

Random
Decompose r = φ · w, V(s) = ψ(s) · w with ${\psi}\; = \;E[{\Sigma}\;{\gamma}^{k}\;{\varphi}]$ successor features; new task = new w, reuse ψ → instant transfer without re-learn; GPI foundation
Same as Q
Not real

Successor features: V = ψ·w; new w reuses ψ for zero-shot task transfer.

#theory

214 interview questions in this topic.