Guillaume Valentis and Quentin Berthelot, ECE Paris Engineering School, France

In order to develop ever more intelligent and autonomous systems, it is necessary to make them self-learning, since it is impossible to include in their program everything they may encounter during their life-cycle. In this research work, we aim at answering the following: if a system’s environment is modified, how could the system respond to it quickly and appropriately enough? We achieve it by using reinforcement learning to allow the system to rate its decisions, then by developing adaptive learning algorithms for gain and loss rewards. The algorithms include probabilities’ analysis providing to the system ability to adapt its knowledge through time and to respond to a changing environment. Simulations are made for a robot finding its exit in a labyrinth. Results show that reinforcement and adaptive learnings can have many useful applications by offering to a system a reliable possibility of evolution within complex environments in specific situations.

Reinforcement Learning, Neural Network, Autonomous Systems, Adaptive Learning, Changing Environment