FLOW-SHOP ROBOTIC SCHEDULING WITH COLLABORATIVE REINFORCEMENT LEARNING

摘要

A collaborative reinforcement learning (RL) method for minimizing make-span in a robotic flow-shop scheduling problem is presented. The robot can operate either autonomously (no adviser) or semiautonomously (with adviser). In autonomous mode, the robot uses RL ε-greedy selection scheme. In semiautonomous mode a collaborative agent (adviser) provides advice to the robot. The robot is endowed with three cognitive abilities: (i) ability to assess its own performance, using an adaptive performance threshold to switch between collaborative modes, (ii) short term ability to assess good and bad advice, and to accept or reject it, (iii) and long term ability to assess advisor's skill levels, and discontinue collaborating with novice advisors. Adviser's behaviors are simulated by various skill levels, represented by soft-max action selection distributions. The collaborative robot-adviser system average error was, at the most, 9.4% within a lower-bound value. An expert adviser was found to accelerate the robot learning process.