Source
Zorowitz, Momennejad & Daw, LIT-1496, read in full in NOTE-1136 from the version of record, Computational Psychiatry 4: 1–17 (2020), CC BY 4.0: pp. 3–9 and 12–14, Figures 1–5.
What was actually shown
Under the optimal Bellman equation, opportunities for reward propagate to earlier states and avoidable threats do not, “because the return-maximizing action is to avoid” (p. 3). Replacing the max over next actions with a mixture of best and worst (weight w) makes the agent expect to fail later. In a gridworld, negative value then spreads from the threat to antecedent states and the optimal path keeps its distance (Fig. 1). The same agent cashes out earlier in the balloon analogue risk task, approaches more slowly as potential loss grows in a sleeping-predator task, and avoids a branch with an early large loss and a later larger gain (Figs 2–3), matching findings in anxious people that the paper cites. In learning, the spread of negative value can end with the agent forgoing action altogether (Fig. 4).
What could have come out otherwise: a pessimistic weight might have produced only local caution near the threat. It produced spread across the whole state space, growing as w falls.
What this does not say
- That anxious people have this misbelief. No behaviour is fitted. The one empirical test the paper proposes, a reduced free-choice premium, is left for future work.
- That the danger is overestimated. The point of the model is that it need not be: the threat’s value and the world’s transitions are correct.
- That avoidance survives correction of the misbelief. The paper argues it “may” (p. 8), by analogy with model-free relearning; it does not simulate it.
- That this is the only placement of the pessimism. Pessimism about the world’s transitions or the current state could, the authors say, produce similar symptoms; only the self-policy version is simulated.
- That it covers all anxiety. The authors exclude OCD compulsions and PTSD memory disturbance (pp. 11–12).
Connections
- CLAIM-446. A rival account of the phenomena the claim explains, in which avoidance is kept by a belief about one’s own future control while the threat is represented correctly. The claim’s test, which holds the represented threat probability fixed, would not separate them; it needs to hold the control belief fixed as well. The paper’s explanation for avoidance outlasting a corrected belief, cached value, is the habit-with-memory alternative the claim’s defeater names.
- THEORY-910. Brown et al. read pessimistic policies as possibly normative under threat rather than as anxiety. This paper allows the same: pessimism “may reflect the exaggerative extremes of an otherwise adaptive evaluation strategy” (p. 11).
- THEORY-285. Both are model-level accounts of a threat representation that maintains avoidance; the maintaining variables differ.