1. Using reinforcement-based models of transitive inference
to simulate primate data
Clara Bergene, Olga F. Lazareva (Drake University), Regina Paxton Gazes (Zoo Atlanta), Robert Hampton (Emory University)
Background
Transitive inference is the ability to deduce that if A > B, and B > C, then A > C
Example of deductive reasoning
In nonverbal tasks, can be explained by appealing to reinforcement history
Reinforcement-based models (Wynne, Siemann-Delius) have been used to accurately predict pigeons’ performance on the tasks
Less is known about the ability of reinforcement-based model to predict primate data
Our goal
Test predictive ability of Siemann-Delius and Wynne models using primate data for seven-item series with backward training and list linking training
Simulation details
Used both Wynne model and Siemann-Delius model (only Siemann-Delius presented)
Selected a combination of free parameters that produced best fit for training data
Used these parameters and accrued associative values to predict testing performance on subject-by-subject basis
List linking training
F+ G-, E+ F-, D+ E-, C+ D-, B+ C-, A+ B-
M+ N-, L+ M-, K+ L-, J+ K-, I+ J-, H+ I-
Linking pair: G+ H-
Tested using within-list pairs (e.g., BD or IM) and between-list pairs (e.g., BI)
Data collected at Hampton’s lab (Emory University)
12 adult rhesus macaques (Macaca mulatta)
Seven-item sequential backward training
F+ G-, E+ F-, D+ E-, C+ D-, B+ C-, A+ B-
Tested using all novel pairs
Training details
Results: List Linking
Captures the difference between end-anchor pairs and inner pairs
Correctly predicts between-list pairs and within-list pairs for List 1 (A…G)
Does not predict within-list pairs for List 2 (H…M)
Possibly due to list linking procedure
Results: Sequential backward training
(averaged across A…G list and H…M list)
Captures difference between the end-anchor and the inner pairs
Predicts opposite to symbolic distance effect
Possibly due to the lack of correction trials or to the backward order of presentation
Sequential backward training with mock correction trials
List linking, mock List 2 training only
Randomly added 1-3 correction trials after each incorrect choice and repeated simulations
Results similar to initial simulations
Run simulations using List 2 training sequence without list linking procedure, and only with initial training
Predicts better performance, but still well below monkeys’ performance
Low performance in List 2 cannot be attributed to list linking procedure alone
Conclusions
Associative models can fit sequential backward training data obtained with pigeons, but not with primates
Possibly due to quicker learning in primates
In list linking, the models provide a good fit for between-list pairs that were assumed to be challenging
They do not, however, predict performance in within-list pairs for List 2 (H…M)
This failure cannot be attributed solely to list linking procedure
Pattern of results suggests contribution from associative values augmented by another process
Sequential backward training in mock forward order
Coded the data in reversed order (F+ G- as A+ B-, and so on) and repeated simulations
Does not predict any difference between end-anchor pairs and inner pairs
Correctly predicts the direction of symbolic distance effect
Backward order of presentation in primates appears to provide a good control for reinforcement history
Acknowledgments
This research was supported by the A&S Drake University Faculty Development Research grant to OFL