Teaching a new dog old tricks - the Assumption Architecture

A rule-based training system for simple LEGO robots

 

 

Abstract:
We present a learning system intended for the simplistic ecological niche of LEGO robots. Binary feedback in the form of good! and bad! from a trainer is enough for the robot to learn a variety of simple tasks. The robot's sensors are structured in groups called senses that pre-process sensor data into useful information. The robot forms consistent assumptions about when to take what action. All feedback is considered significant and may have great impact on the learned assumptions; this is one of the reasons for fast learning. Many experiments were done in simulation, one also with a real LEGO robot.

See the (Word format) paper, accepted for the workshop "Learning'00".