EAGER: Machines that Learn and Teach Seamlessly

点击翻译按钮获取中文摘要

This proposed work seeks to develop a computational approach that can be used to learn a skill from humans and, through the same medium, turn around and teach other less proficient humans what it learned. The learning and teaching will be done through agents that contain the knowledge relevant to the desired skills. Such an agent can be referred to as a learning and teaching agent (LATA). The work plans to accomplish this through two sequential approaches: 1) observational learning (by the agent only), and 2) force feedback learning and teaching. Observational learning will be used to build a minimally proficient LATA agent by observing a human perform the desired task or display the desired skill on a simulator. This agent will be called the baseline agent. This baseline agent will then be enhanced through force feedback learning, where a human will coach the system by providing corrective counter force in real time when the LATA agent errs in its performance of the task. The skills to be learned/taught will require the use of a haptic device such as a joystick or steering wheel as the primary interface. It will learn the actions that the trainer employs to execute the task competently, and use the same haptic device to coach and/or evaluate a less proficient human trainee in learning the same skill. The planned approach centers on using neuroevolutionary techniques. Neuroevolution has been successfully used to address highly complex problems such as pole balancing, abnormal behavior in drivers, and to evolve bots in video games that gradually improve their performance. The proposed work will modify the basic concept of neuroevolutionary techniques as necessary, and apply the resulting system to observational learning as well as force feedback refinement. The testbed domain will be a crane that off-loads boxes or containers from a ship and places them in some other conveyance such as a railroad car or truck. A computer simulation will be used for teaching the LATA agents how to do this. Instructors are increasingly difficult to find, especially for specialty areas that require special skills. From a practical standpoint, this research would give organizations involved in training new tools to teach their constituents. Specific beneficiaries of this technology would include organizations that train students to perform tasks requiring complex motor skills such as driving a car, flying an airplane or operating a crane. The resulting agents could also be used to train operators of tele-operated robots, cranes, unmanned aerial vehicles and other such devices. Surgery training is another potential application of this approach given the appropriate haptic devices. Another interesting application could be for training disabled people basic motor skills.