Addiction as a computational process gone awry.
Level 5 - mechanism / opinion, no new human data
Computational reinforcement-learning model without new empirical human data (mechanism-based reasoning).
PubMed 15591205 · doi:10.1126/science.1102384
What was done
The author developed a computational model of addiction based on temporal-difference reinforcement learning (TDRL). In standard TDRL models of natural learning, actions are reinforced via a dopamine-mediated prediction error signal that is driven toward zero as rewards become fully predicted. The author simulated drug exposure by incorporating a noncompensable, pharmacologically driven dopamine surge into the learning algorithm to observe its effect on action selection.
What was found
The abstract reports no numerical data or quantitative effect sizes. Computationally, adding a noncompensable dopamine increase prevented the error signal from reaching zero, causing the model to continuously reinforce and over-select actions leading to drug receipt.
Why it matters
This work provides a foundational theoretical framework linking the neurobiology of dopamine reward-prediction errors to compulsive drug-seeking behavior through formal reinforcement-learning mathematics.
Limits
The study is a theoretical computational simulation and presents no direct empirical behavioral, animal, or human clinical trial data. The abstract provides no specific mathematical parameters, quantitative validations, or details on biological constraints beyond basic dopamine reward signaling.
Cited by
- supports A 2004 paper by David Redish framed addiction as a computational disease where drugs blocking dopamine reuptake provide dopamine signals that the brain cannot anticipate.