Abstract:
The problem of adaptive human-machine interaction
is investigated. It is sought that not only the human learns how
to perform a task with a novel machine, but the machine itself
co-adapts to the human style in the interaction. This requires
solving the problem of two agents co-adapting or co-learning at
the same time. Due to the lack of human learning and
performance models, it is hypothesized that reinforcement
learning with policy gradient algorithms are good candidates for
addressing this problem with robustness and fast convergence.