Source
Klyubin et al. (2008), LIT-1951; read in NOTE-1514.
What was actually shown
- Actions ±10 N held for 3 steps of 0.02 s; 10 actions; the angle observed in 101 bins; perturbed ensembles of 5⁴ starts; capacity by Blahut to 10⁻⁴ bit.
- With other variables zero, empowerment peaks at upright and drops to zero past a critical angle (Figure 2).
- Over angle and angular speed, the maximum is off-centre, by 0.07 bit, “robust with respect to the variation of parameters”, absent without the cart (p. 6).
- Those states are unreachable “in a sustained run under its own dynamics”, like Garden-of-Eden states (p. 6).
- The greedy controller (Figure 4) “performs classic pole-balancing” (p. 6), with no plan beyond a 1 + 10 step horizon.
What this does not say
- That empowerment maximization generally finds the intuitive task.
- That upright is the empowerment optimum: it is near, not at, the maximum.
- Bearing: CLAIM-354’s reachable set, not a state-wise score, is what separates what an agent can keep from what is merely best placed.