Computational model of reinforcement learning in songbirds
Even in well-learned and precise motor tasks such as reaching or vocalizing, each behavioral rendition deviates slightly from the previous ones. It has been shown that this trial-by-trial variation is not only noise but it can effectively guide adaptive modifications of well-controlled and highly complex motor skills. However, it is currently still unknown how exactly a single rendition shapes and improves future actions. Birdsong learning provides a unique model system to study the neural systems underlying trial-and-error processes of reinforcement learning. We develop linear/nonlinear dynamical system models of reinforcement learning and perform parameter estimation in these models. The model envisioned lives in close symbiosis with experiments. It should raise new questions and suggest further experiments while new experimental results will either validate the model or lead to an extended one – that can then give rise to further questions. Our goal is to estimate crucial mechanistic parameters inherent in reinforcement learning, such as the fraction of behavioral variance accessible for learning. Can such parameters be estimated from behavioral data alone? Additionally, we want to estimate the influence of disrupted versus non-disrupted syllable renditions on future syllables to disentangle their roles in learning.