Richard Watson (2023), Biological Theory 19(1), 22–36.
Key takeaways
From the reading, NOTE-747.
- Comparative, not absolute. A system is “agential (with respect to its parts)” if its organization makes the parts “do what they ‘normally do’ but more, as if the complex constraints between them … were alleviated”. No line between agents and non-agents is drawn.
- Relative to a decomposition. The whole’s norm is borrowed from the parts’ goals. Agency “can be different or absent if we define the agency of the parts differently”.
- The measure. Compare “the total frustration or energy of the system at local equilibria” under the original and the new organization. No formula is given beyond this.
- Goal-directedness is delayed gratification. The system’s dynamics must differ from the objective set by the parts’ constraints, so that it goes uphill “in a specific direction that leads to lower points in the long term”. A ball with momentum is just above zero.
- Organization belongs to the whole’s history. In a two-unit XOR network, correct behaviour leaves each hidden unit free to compute any of 8 functions. The roles settle by mutual symmetry-breaking, and that organization fixes how the network generalizes (Boxes 1, 2).
- It can be learned bottom-up (reported). Hebbian change in a Hopfield-type network, and relationship changes among self-interested agents or species, yield equilibria that better satisfy the original constraints. No system-level teacher or group selection is needed (Watson et al. 2011a, c; Davies et al. 2011; Power et al. 2015).
Standing in the record
Held for QUESTION-009 and read for it. It offers a criterion for when coordination makes a whole more agential than its parts: when the coordination resolves more of the parts’ frustrations than they resolve alone. It refuses to say when a further agent exists outright. So it answers a comparative version of QUESTION-009 rather than the question as asked. Its criterion is filed as THEORY-232. NOTE-747 sets out how it bears on QUESTION-032:
- it agrees with CLAIM-170’s level-relativity;
- it moves the circularity of CLAIM-168 down a level by presupposing the parts’ goals;
- it gives CLAIM-174’s goal-directedness test a behavioural mark;
- it accepts a minimal agency in a ball with momentum, so meets CLAIM-010 by biting the bullet;
- its reported timing result (Tissot et al.) is close to CLAIM-172’s collective variable.
No instruction for machine-learning practice; nothing here belongs in the anthology.