Current artificial intelligence (AI) algorithms for learning control lack time awareness beyond the chronology of events. Time-aware AI can predict the effect of time on the environment and the agent itself, it can understand the predictability based on time, and it can exploit the temporal patterns to reduce its sampling and operating frequency when required. This thesis will show that time awareness in biology is enabled by multiple different systems from enabling slow or fast reactions to measuring passage of time from internal time-keeping oscillatory circuits. Taking inspiration from biology, this thesis will present different models for time-aware control that demonstrate benefits beyond performance in the form of energy efficiency, latency robustness, resilience to input loss, and faster learning due to efficient exploration. Finally, the thesis will demonstrate that to be practical, the models of control need a higher order of adaptivity that can vary its goal/focus on efficiency, performance, and exploration on the fly as the environment and its policy change.
Advisor: Hava Siegelmann