Looking Back and Ahead: Adaptation and Planning by Gradient Descent

摘要

Adaptation and planning are crucial for both biological and artificial agents. In this study, we treat these as an inference problem that we solve using a gradient-based optimization approach. We propose adaptation and planning by gradient descent (APGraDe), a gradient-based computational framework with a hierarchical recurrent neural network (RNN) for adaptation and planning. This framework computes (counterfactual) prediction errors by looking back on past situations based on actual observations and by looking ahead to future situations based on preferred observations (or goal). The internal state of the higher level of the RNN is optimized in the direction of minimizing these errors. The errors for the past contribute to the adaptation while errors for the future contribute to the planning. The proposed APGraDe framework is implemented in a humanoid robot and the robot performs a ball manipulation task with a human experimenter. Experimental results show that given a particular preference, the robot can adapt to unexpected situations while pursuing its own preference through the planning of future actions.