Machine Learning

Behavior Based Robotics is not your mother's AI. Artificial Intelligence v1.0 was basically neither. Early attempts at machine intelligence were from the top-down,  trying to divine symbolic structures that could be used for intelligent inference. This worked for Expert Systems that draw conclusions from a large knowledge base when given a set of related inputs. But navigating the real-world is a much more 'complex' problem. In the 1980's various researchers started on a bottom-up program where intelligence is directly grounded in experience of the world. This is known as Behavior Based. In our system we start with simple sensors and actuators and try to develop a sense of 'place' in the robot.

At the lowest level we use a Stochastic State Machine as a controller. A State Machine is a way of sequencing actions as responses to inputs, and a Stochastic version uses probabilities, rather than fixed parameters, to select the responses. The simplest implementation of this is to have the robot learn how to connect signals from obstacle sensors -- bumpers -- to its drive motors-- avoidance motions. Given a small set of instincts for what behavior is desired -- pleasurable rather than painful -- the robot can learn how to respond to various sensor inputs.

In a set of experiments we calibrated a robot's capabilities and then compared its response when learning how to behave in a simple environment. The calibration involves measuring the behavior when responding in random and ideal manners. A similar measurement then shows that the learned behavior changes from random to something like ideal over time.