Association for Behavior Analysis International

This symposium explores the use of advanced computational and analytical methods in understanding and predicting complex behavioral phenomena. Presentations will examine how machine learning and nonlinear dynamics can inform both basic behavioral processes and real-world therapeutic outcomes. By using these tools, the speakers aim to address key theoretical and practical challenges in behavior analysis. The presentations will discuss how machine learning and nonlinear dynamics enhance our understanding of real-time associative conditioning processes, uncover hidden patterns in therapeutic interactions, provide new perspectives on behavioral allocation, and reveal complex response patterns in probabilistic schedules of reinforcement. Together, these studies will highlight the potential of computational models to refine existing theories and generate new hypotheses that are more in line with the complexities of real-world behavior. By bridging basic experimental research with practical real-world applications, this symposium aims to expand the scope of what can be measured, predicted, and modified within behavior analysis.

This presentation discusses a real-time machine reinforcement learning model of classical conditioning, originally described by Sutton and Barto (1987) and Ludvig, Sutton, and Kehoe (2012). While discussed in computer science and neuroscience, this promising model is often overlooked in behavior analysis. This model uses a temporal difference algorithm to calculate prediction error, which considers differences in prediction across time rather than within a single discrete trial. Like the Rescorla-Wagner model, prediction error is then used to update expected US values. Temporal relationships between stimuli are captured by allowing some trace of a CS to persist after it has ended. Together, these features enable the model to explain a wide range of temporal conditioning phenomena in addition to those explained by the Rescorla-Wagner model. The presentation will explore the model’s foundations in reinforcement learning, its relationship to other classical conditioning models, its strengths and limitations, and its applications in basic research and teaching within behavior analysis.

The matching law states that relative rate of responding equals or “matches” relative rate of reinforcement (Herrnstein, 1961). Herrnstein and Vaughan (1980) proposed melioration as a mechanism for matching. According to melioration, behavior will become allocated to the response alternative with the highest immediate rate of reinforcement. Herrnstein and Vaughan (1980) tested this account against a maximization account which proposes that behavior will become allocated so as to maximize overall rate of reinforcement. Melioration was found to be a better account of matching than maximization. However, according to the quantitative approach taken by machine-based reinforcement learning, melioration and maximization processes should converge to the same results. The discrepancy in empirical results may be due to the extent to which organisms observe environmental states or are sensitive to certain information, a possibility that has been explored empirically. In this talk, we will discuss how the presence or salience of environmental information relates to the melioration and maximization accounts of matching. This will be accomplished through a brief review of empirical work as well as machine-based reinforcement simulations.

Molecular analyses of behavior have been an important concept in behavior analysis since at least the 1930s. Such work has led to important concepts, such as shaping and chaining, which are heavily used in applied settings to teach new skills; bout analyses that allow us to distinguish motoric from motivational aspects of responding; atomic repertoires that form the building blocks of more complex behavior; and the central role of timing in various conditioning paradigms. Historically, analyses of molecular processes of behavior have been conceptual or involved building theoretical models that are tested via experimental preparations. This presentation discusses recent efforts to apply machine learning methods to analyze previously published data from nonhuman organisms responding to various probabilistic schedules of reinforcement. Our findings demonstrate how machine learning can reveal nonlinear patterns in responding that take into account the temporal structure of events and offer new perspectives that enhance existing theoretical frameworks.

This study explores the role of nonlinear dynamics, specifically nonverbal synchrony, in predicting treatment outcomes in a psychiatric inpatient setting. Nonverbal synchrony, defined as the coordinated movement between individuals engaged in information exchange, has been shown to influence psychological processes, including treatment success. The current research employed mediation analysis to investigate how technician experience (in years) impacts treatment outcomes for 61 patients with severe behavior problems (SBP). Both patients and technicians wore biosensors to track motion, enabling the measurement of synchrony. Results revealed that technician experience had a significant direct effect on treatment outcomes (Level 1: ß=0.23, p<.001; Level 2: ß=0.28, p=.024). Moreover, nonverbal synchrony mediated 28% of the effect for patients with Level 1 symptom severity and 18% for those with Level 2 severity. These findings highlight the significance of nonlinear dynamics, such as synchrony, in shaping treatment outcomes and suggest that interpersonal coordination may be a key mechanism by which technician experience enhances therapeutic efficacy.