Applying hybrid hyperdimensional computing to speed imitation learning on the edge

Maintaining an advantage in drone warfare capability is critical to the present and future of national security. This advantage can only be built and maintained with adaptable drones that can update their behavior by incorporating human resourcefulness in minutes on edge hardware. This research makes it possible with current low-cost, low-SWAP-C drone technologies. Countless lives of military personnel can be saved with fast-adapting drone technology.

Small, cheap, low-power computational devices like those found in drones are becoming ubiquitous in modern life. Algorithms are increasingly judged not just on their capabilities but also on how well they run on mobile, resource-constrained devices.

One general use case involves quickly training an automated agent to mimic human operator behavior in an edge-computation environment. Without access to sophisticated reinforcement learning environments or high-performance computing devices required for deep neural networks, the agent would need to learn to mimic human behavior quickly from limited examples.

By incorporating dynamic vector lengths and a vector stacking operation into an imitation learning algorithm, this work showed how new behaviors can be added to drone capabilities within seconds. This is critical to the fast-paced, ever-changing world of drone combat, where adaptability is vital.

This work explored a hybrid hyperdimensional computing neural network imitation learning algorithm that can be used quickly in resource-constrained environments. The capability of the algorithm was demonstrated by duplicating expert behavior in two industry-standard reinforcement learning environments. After watching only one short flight, it also learned to duplicate a trained AI agent in a flight simulation environment.