Intrinsically motivated anticipatory learning utilizing transformation invariance

In this paper, novel reinforcement learning framework with intrinsic motivation to reproduce past successful experience is presented. Geometric transformation invariance is utilized to measure the reproducibility of experience. Top-down “expectation” to reproducibility of experience effectively biases strategy of exploration. As a result of consistent exploration via reproduction of successful experience, learning process is accelerated. Simulation experiments in grid world demonstrate useful characteristics of proposed framework.