Online deep learning for behavior prediction
Anthony M. Bloch · 2022
We implement online deep learning for target behavior prediction. Our online deep learning algorithm provides an autonomous agent the ability to train in real-time while also shaping the frequency of training based on its current performance level. The benefits of our algorithm are twofold: (1) to enable an autonomous agent to train in real-time and continue to learn to accurately predict target behavior even while its target changes the strategy guiding its behavior, and (2) to achieve more efficient usage of its computational resources by managing its training frequency. This trained predictive capability is leveraged in autonomous decision-making to influence a target’s behavior by selecting those actions that produce a predicted response from the target that supports the end goal of the autonomous agent. In our scenario, the goal of the autonomous agent is to influence its target to circle the perimeter of the environment. We test our online deep learning algorithm in environments of varying sizes to demonstrate that the time it takes for an autonomous agent to achieve the target level of accuracy is directly proportional to the size of the environment.