peer-reviewed work

Publications

Academic papers I've contributed to. Click the blog link on any paper to read a plain-English writeup of what we actually did and why it mattered.

01
Conference on Data Science & Management of Data (CODS-COMAD) 2024
Reinforcement Learning ML

Cost-Sensitive Trees for Interpretable Reinforcement Learning

Siddharth Nishtala, Balaraman Ravindran

When a decision tree learns to mimic a neural network's behavior, it treats all mistakes equally — but in reinforcement learning, some wrong actions are far more costly than others. We introduce cost-sensitive variants of two established methods that use the neural network's Q-function to assign different penalties to different errors during tree construction, pushing the tree to avoid the expensive mistakes even at the cost of making more cheap ones. The result: interpretable tree policies that match or beat the originals across four environments, often with shallower trees.

paper pdf blog post
02
Autonomous Agents for Social Good Workshop at AAMAS 2021
Reinforcement Learning ML

Selective Intervention Planning using Restless Multi-Armed Bandits to Improve Maternal and Child Health Outcomes

Siddharth Nishtala, Lovish Madaan, Aditya Mate, Harshavardhan Kamarthi, Anirudh Grama, Divy Thakkar, Dhyanesh Narayanan, Suresh Chaudhary, Neha Madhiwalla, Ramesh Padmanabhan, Aparna Hegde, Pradeep Varakantham, Balaraman Ravindran, Milind Tambe

Knowing which women are at risk of dropping out of a health program is only half the problem — an NGO with three health workers can't personally call everyone the model flags. We formulate this as a Restless Multi-Armed Bandit problem, where each woman is an "arm" with her own re-engagement dynamics, and a limited daily call budget must be allocated to maximize total engagement across the program. Backed by a real randomized pilot study showing personal calls produce a 61% improvement in engagement over no intervention, the system learns to prioritize the women most likely to actually respond — not just the ones most likely to be at risk.

paper pdf blog post
03
Harvard AI for Social Good Workshop 2020
ML

Missed calls, Automated Calls and Health Support: Using AI to Improve Maternal Health Outcomes by Increasing Program Engagement

Siddharth Nishtala, Harshavardhan Kamarthi, Divy Thakkar, Dhyanesh Narayanan, Anirudh Grama, Aparna Hegde, Ramesh Padmanabhan, Neha Madhiwalla, Suresh Chaudhary, Balaraman Ravindran, Milind Tambe

India accounts for 12% of global maternal deaths, and a big part of the problem is information never reaching the women who need it — partly because they quietly disengage from the health programs meant to help them. We built deep learning models trained on 70 million call records from 300,000 women to predict, early, who is at risk of dropping out of ARMMAN's mMitra program — giving health workers a ranked list of who to call before it's too late. The models were validated in a real pilot deployment, not just on a test set.

paper pdf blog post