Discovery of Predictive Representations With a Network of General Value Functions

The ability of an agent to {\em discover} its own learning objectives has long been considered a key ingredient for artificial general intelligence. Breakthroughs in autonomous decision making and reinforcement learning have primarily been in domains where the agent's goal is outlined and clear: such as playing a game to win, or driving safely. Several studies have demonstrated that learning extramural sub-tasks and auxiliary predictions can improve (1) single human-specified task learning, (2) transfer of learning, (3) and the agent's learned representation of the world. In all these examples, the agent was instructed what to learn about. We investigate a framework for discovery: curating a large collection of predictions, which are used to construct the agent's representation of the world. Specifically, our system maintains a large collection of predictions, continually pruning and replacing predictions. We highlight the importance of considering stability rather than convergence for such a system, and develop an adaptive, regularized algorithm towards that aim. We provide several experiments in computational micro-worlds demonstrating that this simple approach can be effective for discovering useful predictions autonomously.