Herself’s Artificial Intelligence

State Machines

A state machine keeps a temporal representation of the state of its environment. It tracks the state it begins in, as well as states in which actions or sensory data change the state.

Finite state machines are used as part of the AI engine in games. This type of setup can be computationally efficient for updating game characters.

In a neural net state machine the states are generally stored in some type of a vector. The network has a hidden unit for each feature of the environment it needs. The input consists of an input for each sense, plus the input from the hidden units. There is an out put for each possible action that can be taken.

Another way to store the information is with a map that is updated. This is known as ‘iconic representation’ and is used for agents.

In a robot an ‘artificial potential field’ is used. The field is a grid with attractive and repulsive potentials summed for each box. The attractive spot is where the robot is supposed to be heading. It’s number is some constant times the distance the robot is from the goal. A repulsive field is an obstacle for the robot to avoid. It is scored as the constant divided by the distance. The motion of the robot is along the gradient, or down the hill to its goal.