Contribution
It recasts empowerment (LIT-1965) as an interventional quantity, adds contextual empowerment with “objective” (state) and “subjective” (sensor history) contexts and a constructible context automaton, and tests the idea in three scenarios: cart-pole, sensor evolution, and a hardware robot.
Key insight
“Empowerment quantifies (1) a potential rather than an actual information flow in the system … (2) it looks at information that could be generated specifically by the agent itself” (p. 11). It is a property of the embodiment, not of any controller: “it does not depend on a particular action-selection mechanism” (p. 11).
Assumptions
- The perception–action loop as a causal Bayesian network, arrows read “as actual causal mechanisms” (p. 2); actions chosen through an independent source Z_t, so information found downstream must have come from it.
- n-step empowerment (eq. 1, p. 3): ℰ(Aⁿ_t → S_{t+n}) = max_{p(a_t, …, a_{t+n−1})} I(A_t, …, A_{t+n−1}; S_{t+n}), open-loop (“the actuation distribution is selected ‘blindly’”); feedback variants are possible “though the computation becomes more intricate”.
- Contextual empowerment (eq. 2): ℰ(Aⁿ_t → S_{t+n} | C) = Σ_c p(c) ℰ(Aⁿ_t → S_{t+n} | C = c), C a non-descendant of the actions. “This can be interpreted as the channel capacity with side information known to the sender … and the receiver.”
- Finite variables; capacity by Blahut’s algorithm.
Key results
- Eq. (3), p. 4: contextual ≥ context-free, since the context-free case is the one that ignores the side information. (Re-derived: for actions independent of C, I(A; S | C) = I(A; S, C) ≥ I(A; S), so Σ_c p(c) max I(A; S | C = c) ≥ max Σ_c p(c) I(A; S | C = c) ≥ max I(A; S).)
- Objective against subjective (p. 3). State-specific empowerment, which “implies access to ‘objective’ state information”, is “never smaller (but in general larger)” than the global value; the agent’s own sensor-action history “constitutes an upper limit on what the agent can possibly know” (p. 4). A context automaton is a finite-memory filter of that history, generalizing ε-machines; its quality is the gain over context-free empowerment, computable without outside variables.
- Cart-pole (pp. 4–7). Two actions ±10 N held for 3τ, 10 actions, the pole’s angle observed in 101 bins; ensembles of 5⁴ perturbed initial states. Empowerment drops to zero past a critical angle (Figure 2). Over angle and angular speed, the maximum is off-centre, 0.07 bit above upright (Figure 3), but those states “are unreachable through the intrinsic dynamics of the agent”, like Garden-of-Eden states. A greedy controller choosing the action whose successor state has the higher 10-step empowerment “performs classic pole-balancing” (Figure 4), a horizon of 1 + 10 steps and “no ‘overarching plan’”.
- Sensor evolution (pp. 7–9). As in 2005: blobs to arcs at about d = 12, read as “analogous to a phase transition” (Figure 5); arced sensors present but inferior below d = 11.
- AIBO (pp. 9–11). Head tilt as action, infra-red distance as sensor, 1,000 actions with a book placed for 100 of every 200. Over |A|, |S| ∈ {2, …, 15}: context-free 0.52–2.03 bits, book-contextual 0.61–2.15, best evolved two-state automaton 0.56–2.15, with fitness ℰ(A → S | C) − I(A; C) to penalize dependence induced by undersampling. Where the book matters, the automaton’s gain approaches the book’s (Figure 7).
- Discussion. Empowerment identifies states from which the agent can best “home into a suddenly emerging goal” (p. 12): from upright the pole can be made to fall right; not conversely. The horizon is “the main free variable”. A weak form of the principle, organisms approximating informationally optimal dynamics, is proposed; a strong one, empowerment computed by biology, is “unlikely”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Contextual empowerment is never lower than context-free | strong | eq. (3); re-derived |
| C2 | Greedy empowerment maximization yields pole balancing with no task | moderate | one simulation, one parameter set, two actions |
| C3 | The highest-empowerment cart-pole states are not reachable under the agent’s own dynamics | moderate | numerical, stated as robust to parameters (p. 6) |
| C4 | Empowerment can drive qualitative sensor change between niches | moderate | one grid world |
| C5 | Empowerment maximization can construct contexts that track external states | moderate | one robot experiment, 1,000 steps, a designed single external factor |
| C6 | Empowerment is a candidate universal utility for biological adaptation | weak | argued |
Concepts
- Context — a variable of non-descendants of the actions used as side information.
- Context automaton — a finite-state machine fed (S_t, A_t) whose state serves as context.
- Open-loop empowerment — capacity with actions chosen without feedback.
Connections
Gibson; Ashby’s requisite variety; homeokinesis and predictive information (Ay et al.); the free energy principle (Friston), compared in the Discussion. Klyubin et al. 2005 (LIT-1965, NOTE-1511) is the context-free origin.
Bearing on the record
- CLAIM-354.
- Reachability against sustainability. The cart-pole shows state-wise empowerment and what the agent can sustain coming apart: the maxima are unreachable, and the agent’s own dynamics settle on upright. This is the claim’s point that the viable alternatives, not states taken one at a time, are what count (CLAIM-079), found inside the empowerment literature.
- A partial order in empowerment’s own terms. Objective ≥ subjective (history) ≥ automaton ≥ context-free is an ordering by what the agent can read, which is CLAIM-354’s note of 2026-11-02 and CLAIM-537’s I ≥ III ≥ II, in bits.
- Not a defeater. Empowerment is a scalar, but it is horizon-dependent (“the main free variable”) and ignores cost and viability constraints unless they are built into the channel, so it does not fix the claim’s other dimensions.
- CLAIM-537. The claim’s three conditions have an information analogue here: the side-information capacity (eq. 2) is the adjustable case with the agent reading the context, context-free is the static case, and the automaton is the observed case with a bounded reader. The correspondence is of structure; the paper has no tolerance model.
- TERM-064. The AIBO result is a constructed legibility: an internal context that comes to carry an external distinction only because the distinction modulates the agent’s influence (“if it does not modulate the influence of the agent on the world, the context change is not considered relevant”, p. 13). That is legibility relative to a task the agent did not choose, the capacity of its own channel.
- No instruction for ML practice (empowerment as an RL intrinsic reward came later and is not in this paper).
Limitations
- Computation needs a “detachable” world model (p. 11); the AIBO experiment was arranged so the causal effect could be read from observation.
- Each scenario is one small model; no comparison against other intrinsic utilities.
- Open-loop throughout.
Open questions
Closed-loop and interleaved empowerment; how a sustainable system’s horizon is set; whether actuator evolution under energetic costs shows the selective pressure the authors predict (p. 13).