S.N.TAKI_
PHD-SEEKING RESEARCHER // TRUSTWORTHY & EXPLAINABLE ML

Shahariar
Newaz Taki

I study why machine learning models fail — and how to make that failure legible. My work spans robustness to distribution shift, uncertainty quantification, and deep learning interpretability, applied to sequence models and multi-agent reinforcement learning for autonomous systems.

VIEW RESEARCHGET IN TOUCH
94.1%
Peak win rate, UCAV DQN
3.02×
Faster MAPPO convergence
322
GRE composite score
Portrait of Shahariar Newaz Taki
DHAKA, BANGLADESH · 23.79°N 90.41°E
SECTOR — 01

Research Interest

What I'm looking to work on next, as a doctoral researcher.

My research focuses on trustworthy and explainable machine learning— model failure characterization, robustness to distribution shifts, uncertainty quantification, and deep learning interpretability. I'm particularly drawn to sequence modeling (LSTMs for time-series forecasting) and representation learning (CNNs for genomic sequence classification), with the broader aim of building AI systems that fail gracefully, reveal their own limitations, and can be audited once deployed in the real world.

CGPA
3.33 / 4.00
GRE
322 (V157 · Q165 · AW4.5)
TOEFL
92 / 120
Location
Dhaka, Bangladesh
SECTOR — 02

Research Experience

Reinforcement learning and sequence modeling for autonomous, multi-agent systems.

UCAV-DDQN

AT-PD³MS-DDQN: Attention-Enhanced Prioritized Dueling Multi-Step Double DQN

UCAV maneuver decision-making in short-range aerial combat

  • Designed a Transformer temporal encoder processing an 8-step historical 6-DOF state window via multi-head self-attention
  • Implemented a dueling architecture decomposing Q-values into state-value and action-advantage streams
  • Developed Hybrid Prioritized Experience Replay (PER-H), weighting sampling by TD-error magnitude and episodic outcome
  • Built an LSTM-based opponent intent prediction module with an automatic curriculum to eliminate cold-start instability
94.1%
win rate vs. maneuvering target
89.7%
win rate vs. adversarial DQN
+10.7pp
over MS-DDQN baseline
37%
faster convergence
HGLM-MAPPO

Hierarchical Graph-Attention LSTM with MAPPO

Autonomous dual-UAV cooperative air combat

  • Designed the HGLM framework: tactical role assigner, LSTM temporal encoder, two-layer graph attention network, and a CTDE MAPPO backbone
  • Developed a four-stage adversarial curriculum self-play training schedule
  • Validated each module via ablations across 7 configurations and 5 random seeds (p < 0.05, paired t-test)
  • Surfaced interpretable inter-agent attention weights that anticipate tactical role-switch events
91.2%
win rate vs. straight-line opponents
82.7%
win rate vs. rule-based adversaries
+37.5pp
over P-DQN baseline
3.02×
faster convergence
MARL-SWARM

Multi-Agent Reinforcement Learning for Autonomous UAV Swarm Control

Coverage, tracking, and formation flying — built from scratch

  • Built three UAV swarm simulation environments from scratch: coverage, tracking, and formation flying
  • Implemented QMIX, MADDPG, and MAPPO in pure NumPy, with a BiLSTM encoder benchmarked against an MLP baseline
  • Introduced the Non-Stationarity Index (NSI), a KL-divergence measure of training instability in multi-agent UAV settings
  • Derived an evidence-based framework for algorithm and architecture selection from 14 empirical findings
TOOLS — NumPy · Python
3
environments implemented
3
MARL algorithms, pure NumPy
14
empirical findings synthesized
NSI
novel instability metric
LSTM-STOCK

Stock Market Prediction using RNN-LSTM

Multi-layer LSTM for financial time-series forecasting

  • Developed a multi-layer LSTM model for time-series forecasting
  • Achieved improved RMSE through feature engineering and optimization
TOOLS — TensorFlow · Keras · PyTorch · Python
NO BENCHMARK METRICS LOGGED
SECTOR — 03

Publications

Manifest, most recent first.

2026

S. N. Taki, “AT-PD³MS-DDQN: Attention-Enhanced Prioritized Dueling Multi-Step Double Deep Q-Network for Autonomous UCAV Maneuver Decision-Making in Short-Range Aerial Combat.”

IN PROGRESS
2026

S. N. Taki, “Hierarchical Graph-Attention LSTM with Multi-Agent Proximal Policy Optimization for Autonomous Dual-UAV Cooperative Air Combat: An Adversarial Curriculum Self-Play Approach.”

IN PREPARATION
2026

S. N. Taki, “Multi-Agent Reinforcement Learning for Autonomous UAV Swarm Control: Implementation, Deep Empirical Analysis, BiLSTM Comparison, and a Novel Research Agenda,” Engineering Applications of Artificial Intelligence (Elsevier).

SUBMITTED
2023

S. Chaki, S. N. Taki, G. H. Siyam, and J. Hasan, “Stock Market Prediction: An Approach Based on Multi-Layer Long Short-Term Memory Network.”

UNDER REVIEW
SECTOR — 04

Work Experience

2025 — Present

IT Executive

Seascape Shipping Line Ltd. (Bangladesh)

Banani, Dhaka-1213, Bangladesh

2022 — 2025

IT Executive

Skama Shipping and Projects Ltd. (Bangladesh)

Banani, Dhaka-1213, Bangladesh

2021 — 2022

Freelance — Web Development

Remote

SECTOR — 05

Education & Test Scores

2016 — 2020

B.Sc. in Computer Science and Engineering

Uttara University, Dhaka 1230, Bangladesh

3.33/ 4.00 CGPA
STANDARDIZED TEST SCORES
322
GRE — Verbal 157 · Quant 165 · AW 4.5
92
TOEFL — L29 · R29 · W19 · S15
SECTOR — 06

Technical Skills

PYTHON — 2+ YEARS
Scikit-learnOpenCVTensorFlowKerasPyTorchMatplotlibNumPyPandas
OTHER LANGUAGES
CC++MySQLUnix Shell Scripting
SOFTWARE
NVivoIBM SPSSLaTeXMicrosoft OfficeAndroid StudioVS CodeCodeBlocksGoogle ColabPyCharmJupyter Notebook
HARDWARE & INSTRUMENTS
ArduinoOscilloscopeTrainer Boards
SECTOR — 07

Presentations, Certifications & Service

ORAL PRESENTATIONS
  • 2021“Machine Learning: Concepts and Applications,” Uttara University, Dhaka
  • 2020“Stock Market Prediction Using Recurrent Neural Network,” Uttara University, Dhaka
  • 2019“Unmanned Aerial Vehicle (UAV) for Transport of Items such as Packages, Medicines, Foods, Postal Mails, and Other Light Goods,” Uttara University, Dhaka
WORKSHOPS & PARTICIPATION
  • Machine Learning and Human-Computer Interaction2019
  • Django and Oracle Based Applications2019
  • Research Methodology and Thesis Writing2018
  • LaTeX2017
  • Object Oriented Programming2016
  • Microsoft Word2016
CERTIFICATIONS
  • Research Methodology: Basic to Advanced — Research Help Bangladesh
  • Developing AI Applications with Python
  • Introduction to DevOps
  • DevOps, Cloud, and Agile Foundations
  • Hands-on Introduction to Linux Commands and Shell Scripting
  • Getting Started with Git and GitHub
  • Sequences, Time Series and Prediction
  • Skill Development for Mobile Game and Application Project — ICT Division (Advanced)
  • Skill Development for Mobile Game and Application Project — ICT Division
LEADERSHIP & MEMBERSHIP
  • Ex-Vice President, Uttara University CSE Student Association — served for the welfare of students at Uttara University.
  • Member, IEEE Student Branch at Uttara University
EXTRA-CURRICULAR
  • Volunteer, IEEE Student Branch Opening Ceremony, Department of Engineering, Uttara University, 2019, Dhaka
  • Volunteer, ACM International Collegiate Programming Contest (ACM-ICPC), 2019, Dhaka
  • Volunteer, ACM International Collegiate Programming Contest (ACM-ICPC), 2017, Dhaka
  • Voluntary Blood Donor since 2018, Quantum Method
SECTOR — 08

References

Dr. MD Mizanur Rahman

Dean, Department of Engineering, Uttara University

Holding 77 Beribadh Road, Dhaka 1230

+880 1778949274[email protected]

Dr. A.H.M. Saifullah Sadi

Chairman, Department of Computer Science, Uttara University

Holding 77 Beribadh Road, Dhaka 1230

+880 1795379956[email protected]