Article Per Year (5 Year)
p-Index From 2021 - 2026
0.23
P-Index
This Author published in this journals
All Journal Indonesian Journal of Machine Learning and Intelligent Systems
Hameed Affifa
Barani Institute of Management Sciences

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

 1 
Memory-Augmented AI for Autonomous UAV Search and Rescue: Deep Reinforcement Learning in Complex, Partially Observable Environments Hina Javed; Hameed Affifa; Imran Ahmad Khan
Indonesian Journal of Machine Learning and Intelligent Systems Vol. 1 No. 1 (2026)
Publisher : Indonesian Artificial Neural Network Society

Show Abstract | Download Original | Original Source | Check in Google Scholar

Abstract

Unknown, partially observable environments The search of autonomous targets remains one of the basic problems of embodied artificial intelligence. We also present RecurrentPPO-LSTM, an agent of memory-based deep reinforcement learning, which not only attains best-in-class state-of-the-art search efficiency on standard Partially Observable Markov Decision Process (POMDP) benchmarks, but also trains 3.4x faster than other memory-augmented agents. As compared to existing schemes that rely on the bidirectional recurrence (orpruing causality) or complex offline pretraining, our agent can be trained using sparse success signals using unidirectional LSTM to guarantee compact belief states. Our implementation on MiniGrid DoorKey-8x8 has 82.4 percent, which corresponds with higher performance than memoryless Proximal Policy Optimization (65.1 percent) and Deep Q-Networks (58.2 percent). We demonstrate that we have 34 steps-to-target reduction and improved resistance to observation noise (sigma = 0.2), arbitrary dropout sporadically (50 percent), and stochastic transitions (slip probability = 0.3). The transfer learning when trained on photorealistic Habitat Matterport 3D has a 68 percent success with 10,000 fine-tuning steps only. It is relatively compared to five stateof-the-art designs (20252026) to confirm that simple causal memory designs are much better than complex bidirectional designs or modular ones when considering real-time autonomous search. We demonstrate in theory that LSTM augmentation causes minimized anticipated regret of finite-memory POMDPs and we can formally ensure empirical excellence.
Co-Authors Hina Javed Imran Ahmad Khan