Discovering Latent States for Model Learning: Applying Sensorimotor Contingencies Theory and Predictive Processing to Model Context
This work addresses the challenge of improving reinforcement learning for robots by reducing the need to learn each skill from scratch, though it appears incremental as it builds on existing cognitive theories.
The paper tackles the problem of enabling autonomous robots to acquire generic knowledge for learning new skills by developing a world model inspired by predictive processing and sensorimotor contingencies theory, and it evaluates this model in simulation experiments.
Autonomous robots need to be able to adapt to unforeseen situations and to acquire new skills through trial and error. Reinforcement learning in principle offers a suitable methodological framework for this kind of autonomous learning. However current computational reinforcement learning agents mostly learn each individual skill entirely from scratch. How can we enable artificial agents, such as robots, to acquire some form of generic knowledge, which they could leverage for the learning of new skills? This paper argues that, like the brain, the cognitive system of artificial agents has to develop a world model to support adaptive behavior and learning. Inspiration is taken from two recent developments in the cognitive science literature: predictive processing theories of cognition, and the sensorimotor contingencies theory of perception. Based on these, a hypothesis is formulated about what the content of information might be that is encoded in an internal world model, and how an agent could autonomously acquire it. A computational model is described to formalize this hypothesis, and is evaluated in a series of simulation experiments.