Storyboard

Forward prediction is essential to overcome neural processing delays (~100200 ms in humans). Without prediction, a control system with significant delay is unstable. The central nervous system maintains internal models of the dynamics of the body and the environment that allow the consequences of actions to be predicted before sensory feedback arrives.

The linear state model (F·x_k) describes how the state of the world is expected to evolve. The prediction error grows with time: the covariance P(t) increases indefinitely if it is not corrected with new observations. The useful prediction horizon t_div is the time after which the uncertainty grows so much that the prediction loses value.

Projectile trajectory prediction (parabolic) is a computation that vertebrates must perform to catch objects in flight (frogs catching flies, baseball players). Experiments suggest that organisms predict the point of interception rather than continuously following the trajectory, consistent with the kinematic model that incorporates gravity as a prior.

The principle of shock copy (shock corollary, von Holst): each motor command simultaneously generates a prediction of its sensory consequences (_reaf = M·u). The actual sensory signal is compared to the prediction: if they match, the movement was expected and is suppressed (explaining why we don't tickle ourselves). If they do not match (e_unexpected 0), something unexpected happened in the environment.

Model-based predictive control (MPC) is the optimal control paradigm that optimizes a cost function over a future horizon N. It allows anticipating constraints (obstacles, force limits) and planning smooth trajectories. The cerebellum implements the forward model component, and the prefrontal cortex implements the long-term planning component of the MPC.

>Model

ID:('ky', 584)

Description

ID:(0, 584)