machine-learning

agents which learn from the sensory data it receives, rather than a person designing it by hand.

instead of mapping percepts to function as in the agent-function we map inputs $x$ to outputs $y$ by approximating a function $y=f(x)$. The true function $y=f(x)$ is unknown. We assume there is a relationship between the data and the labels of the data (in the case of supervised-learning, otherwise we are looking for relationships as in unsupervised-learning). Our approximation is the hypothesis function. $\hat{y}=h(x,w)$.

• $y$ is desired output
• $\hat{y}$ is prediction/output
• $h$ is hypothesis
• $(x,w)$ is input
• $x$ and $w$ are parameters

The approximated function can also be called the model. The accuracy of the approximation is measure by a loss-function