Reinforcement Learning with Intelligent Context-based Exploration (RL-ICE)

To address the fundamental problem of training rare-object detection systems, SoarTech proposes Reinforcement-Learning with Intelligent Contextual Exploration (RL-ICE). With Deep Reinforcement Learning (DRL), RL-ICE trains an adversarial agent to identify weaknesses in existing object classification systems and generate new, targeted synthetic imagery to improve performance. In Phase 1, SoarTech demonstrated using DRL to improve object recognition through the generation of synthetic imagery: (1) designing effective reward functions for DRL, (2) augmenting very small numbers of training imagery including zero-shot learning (e.g. learning with no labeled examples of the target object). In Phase 2, development of RL-ICE is focused on generalizing the algorithms developed in Phase 1, including: (1) conducting formal evaluations across a wide range of objects, (2) identifying the process for generating maximal performance regardless of the object type, and (3) learning to incorporate other data sources to further enhance algorithmic performance. Furthermore, beyond algorithmic generalization, option period development will investigate transitioning to other applications outside of satellite imagery. This includes aerial imagery with higher slant angles and medium-wave infrared imagery (MWIR). For MWIR processing in particular, SoarTech will work sensing experts at the Michigan Technical Research Institute (MTRI) to transition RL-ICE into a new sensing domain.