Reward Prediction-Error Fidelity: a testable hypothesis linking reinforcement learning mechanisms to applied behavior analysis for autism and neurodevelopmental conditions.
- 1SiPGI, Postgraduate School of Integrated Gestalt Psychotherapy, Torre Annunziata, Naples, Italy.
Abstract
Applied behavior analysis (ABA) has developed an effective technology of reinforcement delivery, including Discrete Trial Training, Natural Environment Teaching, and token economies, largely on functional-behavioral grounds. Computational neuroscience has independently produced formal accounts of reinforcement learning centered on prediction errors and credit assignment. I propose a testable hypothesis connecting the two: the effectiveness of a reinforcement procedure depends on the quality of the learning signal that its delivery produces on each occasion, a property I term Reward Prediction-Error Fidelity (RPE Fidelity). The property is decomposed into four dimensions, temporal credit, contingency, prediction, and outcome value, each corresponding to a construct long studied in behavior analysis. Formalized within an actor-critic model with eligibility traces, the four dimensions enter the learning update multiplicatively, which implies that they cannot compensate for one another and that the effect of each depends on the others and on the learner's history. The combination rule distinguishes the hypothesis from the separate behavioral constructs it draws on, and its dependence on learning history distinguishes it from steady-state quantitative models of reinforcement. I situate the hypothesis within an integrative framework, discuss its application to autistic learners, including differences in reward processing and volatility estimation and the role of learner assent, and compare it with procedural fidelity, quantitative behavioral theory, and alternative learning models, including attentional, timescale-invariant, successor-representation, and Bayesian accounts. I derive nine predictions, specifying for each the result that would favor the hypothesis and the result that would favor its strongest competitor, and outline a staged research programme compatible with single-case experimental designs.
Related
- Motor, Postural, and Communicative Phenotype in SHQ1-Related Neurodevelopmental Disorder with Dystonia and Seizures: An Interdisciplinary Case Report and Hypothesis-Generating Physiotherapy Framework.
- Spread of Chronic Wasting Disease under Stochastic Environmental Conditions and its Control using Deep Reinforcement Learning
- Beyond the Microbial Hypothesis: Shared Genetic, Osteoimmune and Vascular Mechanisms in the Oral-Brain Axis Linking Periodontitis and Alzheimer's Disease.
- Cool and warm conditions disrupt development through divergent mechanisms in larvae of the reef-building coral Orbicella faveolata
- Neopep: Reinforcement Learning From Machine-Learned Force Field Energy for Non-Canonical Peptide Structure Prediction
BibTeX
@article{cesarano2026reward,
title = {{Reward Prediction-Error Fidelity: a testable hypothesis linking reinforcement learning mechanisms to applied behavior analysis for autism and neurodevelopmental conditions.}},
author = {Sebastian G D Cesarano},
journal = {Frontiers in behavioral neuroscience},
year = {2026},
month = oct,
doi = {10.3389/fnbeh.2026.1974577},
url = {https://europepmc.org/article/MED/43011253},
note = {CC BY},
}
RIS
TY - JOUR
TI - Reward Prediction-Error Fidelity: a testable hypothesis linking reinforcement learning mechanisms to applied behavior analysis for autism and neurodevelopmental conditions.
AU - Sebastian G D Cesarano
PY - 2026
DA - 2026/10/27
JO - Frontiers in behavioral neuroscience
DO - 10.3389/fnbeh.2026.1974577
UR - https://europepmc.org/article/MED/43011253
AB - Applied behavior analysis (ABA) has developed an effective technology of reinforcement delivery, including Discrete Trial Training, Natural Environment Teaching, and token economies, largely on functional-behavioral grounds. Computational neuroscience has independently produced formal accounts of reinforcement learning centered on prediction errors and credit assignment. I propose a testable hypothesis connecting the two: the effectiveness of a reinforcement procedure depends on the quality of the learning signal that its delivery produces on each occasion, a property I term Reward Prediction-Error Fidelity (RPE Fidelity). The property is decomposed into four dimensions, temporal credit, contingency, prediction, and outcome value, each corresponding to a construct long studied in behavior analysis. Formalized within an actor-critic model with eligibility traces, the four dimensions enter the learning update multiplicatively, which implies that they cannot compensate for one another and that the effect of each depends on the others and on the learner's history. The combination rule distinguishes the hypothesis from the separate behavioral constructs it draws on, and its dependence on learning history distinguishes it from steady-state quantitative models of reinforcement. I situate the hypothesis within an integrative framework, discuss its application to autistic learners, including differences in reward processing and volatility estimation and the role of learner assent, and compare it with procedural fidelity, quantitative behavioral theory, and alternative learning models, including attentional, timescale-invariant, successor-representation, and Bayesian accounts. I derive nine predictions, specifying for each the result that would favor the hypothesis and the result that would favor its strongest competitor, and outline a staged research programme compatible with single-case experimental designs.
SN - 1662-5153
ER -