Which criteria for autonomously shifting between goal-directed and habitual behaviors in robots?

Research in the fields of Psychology and Neuroscience have provided strong evidence that mammals can adaptively switch between goal-directed behaviors - i.e. deliberative decisions based on costly but flexible planned long-term consequences of actions - and habitual behaviors - i.e. reactive behaviors that are efficient when the environment is stable but inflexible in the case of environmental changes. However, the computational principles underlying this switching ability are not yet understood, and several alternative criteria have been proposed, each tested on specific subsets of experimental datasets. Here we present a neurorobotic implementation and comparison of such type of criteria, plus some new ones imported from the field of ensemble reinforcement learning, with a twofold objective: on the one hand exploring the possible efficiency of such bio-inspired principles to enable robots to have more behavioral flexibility during autonomous development and learning; on the other hand, analyzing whether an asynchronous continuous robotic simulation and comparison of these criteria in a common task can feed current debates in the Psychological and Neuroscience fields. We evaluate these methods in an apparently simple repetitive cube-pushing task on a simulated conveyor belt, but which imposes to the robot constant trade-offs between speed and accuracy and between stability and abrupt changes. Our results show that if overall performance is not improved by using multiple behavioral systems in a stable environment, these methods allow for a better adaptation to environmental changes. The Voting methods and Boltzmann addition, from ensemble reinforcement learning, give the best performance, providing an interesting alternative to Expert selection.