Probability matching in perceptrons: Effects of conditional dependence and linear nonseparability

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Probability matching describes an agent's behavior aligning with environmental event likelihoods.
• Artificial neural networks (ANNs) exhibiting probability matching adjust output activity to reflect past reward probabilities.
• Previous work showed simple ANNs (perceptrons) achieve probability matching with isolated cues.

Purpose of the Study:

• To investigate perceptron probability matching with simultaneous, independent cues.
• To examine perceptron performance when cue independence is violated by logical combinations.
• To identify factors predicting perceptron performance in complex cue environments.

Main Methods:

• Simulations using perceptrons presented with up to four simultaneous cues.
• Varying cue independence and introducing logical combinations (AND, XOR) between cues.
• Manipulating reward sizes associated with cue interactions.

Main Results:

• Perceptrons successfully matched reward probabilities with independent simultaneous cues.
• Performance was better predicted by reward magnitude than by the logical structure of cue interactions.
• Perceptrons treated independent cues as separate information sources for reward likelihood.

Conclusions:

• Perceptrons learn to match probabilities by assuming cue independence.
• Quantitative measures of input signal independence are key predictors of perceptron performance.
• The complexity of logical cue interactions does not solely determine perceptron learning success.