Annika Österdiekhoff
Bielefeld University
Humans frequently engage in multitasking. It can be at work, where we check our email while attending a meeting, or while driving, where we change the radio channel while actively steering. We do this even though it is more error-prone, leads to poorer performance, and can lead to dangerous situations. However, on what criteria do humans switch from one task to another in a sequential multitasking setup, and can we use these criteria to train computational agents to maximize performance while task switching? One possible factor influencing task switching is sense of control (SoC), which arises from the mismatch between predicted and actual outcomes. We propose a framework of three reinforcement learning agents, where a higher-level meta-agent decides to switch between two tasks that are individually solved by a lower-level agent. We integrate a sense of control as an additional reward signal to guide the task switching behavior. The meta-agent may draw from the SoC to decide when to switch to individual task agents. How does a sense of control change task switching behavior? How does the framework of RL agents compare to human task switching behavior?
DFG SPP "The Active Self" 