Towards more practical reinforcement learning

The fields of machine learning, mobile robotics and machine vision have grown steadily closer in recent years, to the extent that learning has been suggested as the best means of producing sophisticated controllers for mobile robots. Such an approach may have merit, but only if the structures and mechanisms provided for learning are tuned to the special needs of robots. These needs are outlined, and reinforcement learning is promoted as the best starting point for fulfilling them. In order to make good on the promise of learning to the level required of mobile robots, significant enhancements are required to current formulations of reinforcement learning. The issues involved in making improvements are discussed, and a simple enhanced model of reinforcement learning is suggested as a first step in this direction