Anytime learning was used by Grefenstette to allow a learning component to continually compute a best solution while the robot operated in the environment using the latest best solution. A natural extension of this for simple robots is to have the learning component off-line while the operations component which is on the robot receives periodic downloads of the best solution. Grefensette's system can adapt quickly to changes in the robot by having the robot's sensors continually update the status of the robot's capabilities in the learning element's simulation. This is where a difficulty develops with an off-line learning system. Assuming there is no direct link from the robot's sensors to the learning system, learning needs to take place based on the robot's observed performance. In addition, information concerning the changing capabilities of the robot is also not available to the learning system. All that is available is general information about the success of the robot.
There are two main issues to address concerning the development of a
system for anytime learning in evolutionary robotics; anytime learning
without sensors and interactively connecting simulations to provide
efficient ER training. A solution for both of these issues involves
the use of anytime learning with modifications to compensate for the
lack of internal sensors in our robots. We call this learning
system Punctuated Anytime Learning since there are periodic times
of increased learning. Training with a GA takes place off-line
on a simple simulation. Periodic checks on the actual robot help to
verify the simulation's accuracy. This solution has
two variations: (1) punctuated anytime learning using the co-evolution of
the robot model parameters, (2) punctuated anytime learning with fitness
biasing.