Conclusion
Cooperation of multiple reinforcement learning algorithms

The combination of multiple modules for solving a task in autonomous robotics (and in particular for navigation) has been little explored to date. However, the complementarities in terms of adaptation speed and computational cost of the different types of reinforcement learning algorithms likely to be used, on the one hand, and the neuroscience results demonstrating the use of multiple learning systems in animals, on the other, argue in favor of designing control architectures integrating multiple algorithms.

It's this still-limited exploration of the possibilities of such combinations in robotics that explains why many of the algorithms reviewed here have their origins in computational neuroscience, and their initial aim is to explain biological data. However, these methods, expressed in the formalism of reinforcement learning, are perfectly adaptable and testable...