scholarly journals A Normative Account of Confirmation Bias During Reinforcement Learning

2021 ◽  
pp. 1-31
Author(s):  
Germain Lefebvre ◽  
Christopher Summerfield ◽  
Rafal Bogacz
Keyword(s):  
Reinforcement Learning ◽  
Prediction Errors ◽  
Option Value ◽  
Confirmatory Bias ◽  
Learning Rules ◽  
Reward Prediction ◽  
Wide Range ◽  
The Face ◽  
Paradoxical Finding ◽  
Normative Account

Abstract Reinforcement learning involves updating estimates of the value of states and actions on the basis of experience. Previous work has shown that in humans, reinforcement learning exhibits a confirmatory bias: when the value of a chosen option is being updated, estimates are revised more radically following positive than negative reward prediction errors, but the converse is observed when updating the unchosen option value estimate. Here, we simulate performance on a multi-arm bandit task to examine the consequences of a confirmatory bias for reward harvesting. We report a paradoxical finding: that confirmatory biases allow the agent to maximize reward relative to an unbiased updating rule. This principle holds over a wide range of experimental settings and is most influential when decisions are corrupted by noise. We show that this occurs because on average, confirmatory biases lead to overestimating the value of more valuable bandits and underestimating the value of less valuable bandits, rendering decisions overall more robust in the face of noise. Our results show how apparently suboptimal learning rules can in fact be reward maximizing if decisions are made with finite computational precision.

scholarly journals A normative account of confirmatory biases during reinforcement learning

2020 ◽  
Author(s):  
Germain Lefebvre ◽  
Christopher Summerfield ◽  
Rafal Bogacz
Keyword(s):  
Reinforcement Learning ◽  
Prediction Errors ◽  
Option Value ◽  
Confirmatory Bias ◽  
Reward Prediction ◽  
Wide Range ◽  
The Face ◽  
Paradoxical Finding ◽  
Learning Policies ◽  
Normative Account

AbstractReinforcement learning involves updating estimates of the value of states and actions on the basis of experience. Previous work has shown that in humans, reinforcement learning exhibits a confirmatory bias: when updating the value of a chosen option, estimates are revised more radically following positive than negative reward prediction errors, but the converse is observed when updating the unchosen option value estimate. Here, we simulate performance on a multi-arm bandit task to examine the consequences of a confirmatory bias for reward harvesting. We report a paradoxical finding: that confirmatory biases allow the agent to maximise reward relative to an unbiased updating rule. This principle holds over a wide range of experimental settings and is most influential when decisions are corrupted by noise. We show that this occurs because on average, confirmatory biases overestimate the value of more valuable bandits, and underestimate the value of less valuable bandits, rendering decisions overall more robust in the face of noise. Our results show how apparently suboptimal learning policies can in fact be reward-maximising if decisions are made with finite computational precision.

scholarly journals A nonlinear relationship between prediction errors and learning rates in human reinforcement learning

2019 ◽  
Author(s):  
Erdem Pulcu
Keyword(s):  
Reinforcement Learning ◽  
Nonlinear Relationship ◽  
Prediction Errors ◽  
Learning Rates ◽  
The Face ◽  
In The Wild ◽  
Actual Outcome ◽  
Update Rules ◽  
Different Sources

AbstractWe are living in a dynamic world in which stochastic relationships between cues and outcome events create different sources of uncertainty1 (e.g. the fact that not all grey clouds bring rain). Living in an uncertain world continuously probes learning systems in the brain, guiding agents to make better decisions. This is a type of value-based decision-making which is very important for survival in the wild and long-term evolutionary fitness. Consequently, reinforcement learning (RL) models describing cognitive/computational processes underlying learning-based adaptations have been pivotal in behavioural2,3 and neural sciences4–6, as well as machine learning7,8. This paper demonstrates the suitability of novel update rules for RL, based on a nonlinear relationship between prediction errors (i.e. difference between the agent’s expectation and the actual outcome) and learning rates (i.e. a coefficient with which agents update their beliefs about the environment), that can account for learning-based adaptations in the face of environmental uncertainty. These models illustrate how learners can flexibly adapt to dynamically changing environments.

scholarly journals Dopamine transients delivered in learning contexts do not act as model-free prediction errors

2019 ◽  
Author(s):  
Melissa J. Sharpe ◽  
Hannah M. Batchelor ◽  
Lauren E. Mueller ◽  
Chun Yun Chang ◽  
Etienne J.P. Maes ◽  
...  
Keyword(s):  
Reinforcement Learning ◽  
Associative Learning ◽  
Prediction Error ◽  
Error Term ◽  
Neural Correlates ◽  
Dopamine Neurons ◽  
Prediction Errors ◽  
Model Free ◽  
Reward Prediction ◽  
Excess Value

AbstractDopamine neurons fire transiently in response to unexpected rewards. These neural correlates are proposed to signal the reward prediction error described in model-free reinforcement learning algorithms. This error term represents the unpredicted or ‘excess’ value of the rewarding event. In model-free reinforcement learning, this value is then stored as part of the learned value of any antecedent cues, contexts or events, making them intrinsically valuable, independent of the specific rewarding event that caused the prediction error. In support of equivalence between dopamine transients and this model-free error term, proponents cite causal optogenetic studies showing that artificially induced dopamine transients cause lasting changes in behavior. Yet none of these studies directly demonstrate the presence of cached value under conditions appropriate for associative learning. To address this gap in our knowledge, we conducted three studies where we optogenetically activated dopamine neurons while rats were learning associative relationships, both with and without reward. In each experiment, the antecedent cues failed to acquired value and instead entered into value-independent associative relationships with the other cues or rewards. These results show that dopamine transients, constrained within appropriate learning situations, support valueless associative learning.

scholarly journals From internal models toward metacognitive AI

2021 ◽  
Author(s):  
Mitsuo Kawato ◽  
Aurelio Cortese
Keyword(s):  
Reinforcement Learning ◽  
Computational Models ◽  
Monitoring Network ◽  
Internal Models ◽  
Inverse Model ◽  
Small Samples ◽  
Prediction Errors ◽  
Hierarchical Reinforcement Learning ◽  
Reward Prediction ◽  
Higher Cognitive Functions

AbstractIn several papers published in Biological Cybernetics in the 1980s and 1990s, Kawato and colleagues proposed computational models explaining how internal models are acquired in the cerebellum. These models were later supported by neurophysiological experiments using monkeys and neuroimaging experiments involving humans. These early studies influenced neuroscience from basic, sensory-motor control to higher cognitive functions. One of the most perplexing enigmas related to internal models is to understand the neural mechanisms that enable animals to learn large-dimensional problems with so few trials. Consciousness and metacognition—the ability to monitor one’s own thoughts, may be part of the solution to this enigma. Based on literature reviews of the past 20 years, here we propose a computational neuroscience model of metacognition. The model comprises a modular hierarchical reinforcement-learning architecture of parallel and layered, generative-inverse model pairs. In the prefrontal cortex, a distributed executive network called the “cognitive reality monitoring network” (CRMN) orchestrates conscious involvement of generative-inverse model pairs in perception and action. Based on mismatches between computations by generative and inverse models, as well as reward prediction errors, CRMN computes a “responsibility signal” that gates selection and learning of pairs in perception, action, and reinforcement learning. A high responsibility signal is given to the pairs that best capture the external world, that are competent in movements (small mismatch), and that are capable of reinforcement learning (small reward-prediction error). CRMN selects pairs with higher responsibility signals as objects of metacognition, and consciousness is determined by the entropy of responsibility signals across all pairs. This model could lead to new-generation AI, which exhibits metacognition, consciousness, dimension reduction, selection of modules and corresponding representations, and learning from small samples. It may also lead to the development of a new scientific paradigm that enables the causal study of consciousness by combining CRMN and decoded neurofeedback.

scholarly journals The modulation of savouring by prediction error and its effects on choice

eLife ◽  
2016 ◽  
Vol 5 ◽  
Author(s):  
Kiyohito Iigaya ◽  
Giles W Story ◽  
Zeb Kurth-Nelson ◽  
Raymond J Dolan ◽  
Peter Dayan
Keyword(s):  
Prediction Errors ◽  
Risk Seeking ◽  
Reward Prediction ◽  
Behavioral Studies ◽  
Wide Range ◽  
Advance Information ◽  
Uncertain Future ◽  
Future Outcomes ◽  
Reward Information ◽  
Reinforcement Learning Model

When people anticipate uncertain future outcomes, they often prefer to know their fate in advance. Inspired by an idea in behavioral economics that the anticipation of rewards is itself attractive, we hypothesized that this preference of advance information arises because reward prediction errors carried by such information can boost the level of anticipation. We designed new empirical behavioral studies to test this proposal, and confirmed that subjects preferred advance reward information more strongly when they had to wait for rewards for a longer time. We formulated our proposal in a reinforcement-learning model, and we showed that our model could account for a wide range of existing neuronal and behavioral data, without appealing to ambiguous notions such as an explicit value for information. We suggest that such boosted anticipation significantly drives risk-seeking behaviors, most pertinently in gambling.

How We Learn to Make Decisions: Rapid Propagation of Reinforcement Learning Prediction Errors in Humans

2014 ◽  
Vol 26 (3) ◽  
pp. 635-644 ◽  
Author(s):  
Olav E. Krigolson ◽  
Cameron D. Hassall ◽  
Todd C. Handy
Keyword(s):  
Reinforcement Learning ◽  
Prediction Error ◽  
Human Error ◽  
Dopamine Neurons ◽  
Prediction Errors ◽  
Neural Basis ◽  
Error Related Negativity ◽  
Reward Positivity ◽  
Reward Prediction ◽  
Feedback Error

Our ability to make decisions is predicated upon our knowledge of the outcomes of the actions available to us. Reinforcement learning theory posits that actions followed by a reward or punishment acquire value through the computation of prediction errors—discrepancies between the predicted and the actual reward. A multitude of neuroimaging studies have demonstrated that rewards and punishments evoke neural responses that appear to reflect reinforcement learning prediction errors [e.g., Krigolson, O. E., Pierce, L. J., Holroyd, C. B., & Tanaka, J. W. Learning to become an expert: Reinforcement learning and the acquisition of perceptual expertise. Journal of Cognitive Neuroscience, 21, 1833–1840, 2009; Bayer, H. M., & Glimcher, P. W. Midbrain dopamine neurons encode a quantitative reward prediction error signal. Neuron, 47, 129–141, 2005; O'Doherty, J. P. Reward representations and reward-related learning in the human brain: Insights from neuroimaging. Current Opinion in Neurobiology, 14, 769–776, 2004; Holroyd, C. B., & Coles, M. G. H. The neural basis of human error processing: Reinforcement learning, dopamine, and the error-related negativity. Psychological Review, 109, 679–709, 2002]. Here, we used the brain ERP technique to demonstrate that not only do rewards elicit a neural response akin to a prediction error but also that this signal rapidly diminished and propagated to the time of choice presentation with learning. Specifically, in a simple, learnable gambling task, we show that novel rewards elicited a feedback error-related negativity that rapidly decreased in amplitude with learning. Furthermore, we demonstrate the existence of a reward positivity at choice presentation, a previously unreported ERP component that has a similar timing and topography as the feedback error-related negativity that increased in amplitude with learning. The pattern of results we observed mirrored the output of a computational model that we implemented to compute reward prediction errors and the changes in amplitude of these prediction errors at the time of choice presentation and reward delivery. Our results provide further support that the computations that underlie human learning and decision-making follow reinforcement learning principles.

scholarly journals Pupil responses as indicators of value-based decision-making

2018 ◽  
Author(s):  
Joanne C. Van Slooten ◽  
Sara Jahfari ◽  
Tomas Knapen ◽  
Jan Theeuwes
Keyword(s):  
Decision Making ◽  
Reinforcement Learning ◽  
Computational Model ◽  
Cognitive Processes ◽  
Learning Task ◽  
Brain Regions ◽  
Computational Approach ◽  
Prediction Errors ◽  
Reward Prediction ◽  
Exciting Possibility

AbstractPupil responses have been used to track cognitive processes during decision-making. Studies have shown that in these cases the pupil reflects the joint activation of many cortical and subcortical brain regions, also those traditionally implicated in value-based learning. However, how the pupil tracks value-based decisions and reinforcement learning is unknown. We combined a reinforcement learning task with a computational model to study pupil responses during value-based decisions, and decision evaluations. We found that the pupil closely tracks reinforcement learning both across trials and participants. Prior to choice, the pupil dilated as a function of trial-by-trial fluctuations in value beliefs. After feedback, early dilation scaled with value uncertainty, whereas later constriction scaled with reward prediction errors. Our computational approach systematically implicates the pupil in value-based decisions, and the subsequent processing of violated value beliefs, ttese dissociable influences provide an exciting possibility to non-invasively study ongoing reinforcement learning in the pupil.

scholarly journals Rethinking dopamine as generalized prediction error

2017 ◽  
Author(s):  
Matthew P.H. Gardner ◽  
Geoffrey Schoenbaum ◽  
Samuel J. Gershman
Keyword(s):  
Reinforcement Learning ◽  
Prediction Error ◽  
Dopamine Neurons ◽  
Prediction Errors ◽  
Model Free ◽  
Reward Prediction ◽  
Midbrain Dopamine ◽  
Sensory Prediction ◽  
Lines Of Evidence ◽  
Midbrain Dopamine Neurons

AbstractMidbrain dopamine neurons are commonly thought to report a reward prediction error, as hypothesized by reinforcement learning theory. While this theory has been highly successful, several lines of evidence suggest that dopamine activity also encodes sensory prediction errors unrelated to reward. Here we develop a new theory of dopamine function that embraces a broader conceptualization of prediction errors. By signaling errors in both sensory and reward predictions, dopamine supports a form of reinforcement learning that lies between model-based and model-free algorithms. This account remains consistent with current canon regarding the correspondence between dopamine transients and reward prediction errors, while also accounting for new data suggesting a role for these signals in phenomena such as sensory preconditioning and identity unblocking, which ostensibly draw upon knowledge beyond reward predictions.

Sign in / Sign up

Export Citation Format

Share Document