T-Lind's Research

REPNET (2023)

Designed and implemented the REProductive NETwork (REPNET) architecture, a custom data structure that reduces reinforcement learning training times by 29.3% and variability by 72.62%. REPNET dynamically prunes unproductive model branches using Adaptive Pruning Time Adjustment (APTA) and non-linear reproduction thresholds, significantly accelerating training in environments like OpenAI’s "CartPole." This innovation optimizes AI model training and has applications across healthcare and sustainability technologies. REPNET is patent-pending.

Cost-based Passive Observation of Events (2024)

This paper presents a cost-based approach to planning for passive observation of unpredictable events, building on previous work that used a Markov model. The authors introduce a cost matrix to account for real-world factors like distance traveled or resources expended, aiming to minimize costs between states rather than just the number of steps. They demonstrate the feasibility of calculating an optimal policy using this cost-based modeling approach through experiments on a "Wedding FOM" scenario. The results show that while the cost-optimized algorithm doesn't significantly reduce the number of steps compared to the no-cost algorithm, it does achieve a statistically significant reduction in overall costs. The authors also developed a Python library and JSON configuration system to make their approach more accessible to other researchers and practitioners. The source code can be found here.