Ashish was a postdoctoral researcher at the Automatic Control Laboratory, ETH Zurich, in 2018.

He is interested in the mathematical underpinnings of decision-making in dynamic and uncertain environments, specifically in optimization, control, game theory, behavioural economics, networks and stochastic systems.

INDIAai interviewed Ashish to get his perspective on AI.

Tell me about your interest in smart energy systems, automotive, and the control of epidemics on networks.

I am broadly interested in problems with decision-making in uncertain and dynamic environments. Research problems of this nature arise in many applications, including energy systems, autonomous driving and control of epidemics over networks. 

A growing share of renewable energy and energy-aware smart buildings pose significant challenges in the operation of the electricity grid in modern times. Electricity system operators solve a large-scale optimization problem to determine the optimal dispatch of power-generating units in a day-ahead manner. However, since renewable energy generation is highly uncertain and hard to predict and regulate, it makes scheduling conventional power plants challenging. Together with a few collaborators, we have recently proposed stochastic optimization techniques that exploit past data of renewable energy generation to efficiently compute generation schedules of conventional power plants that provide precise statistical guarantees on the risk of operational constraint violation. Our current focus is on delivering ancillary service support to the grid by exploiting the flexibility available at distribution networks due to growth in rooftop solar generation, battery energy storage systems and electric vehicles. 

Similarly, my group has developed risk-sensitive optimal control techniques for collision avoidance in multi-robot systems and autonomous driving applications. Our work on this topic was recently awarded the most innovative project in the UG Category by AI and Technology Research Park (ARTPARK), established at IISc, Bangalore, in their innovation grant program and is set to appear at the top robotics conference ICRA 2023. 

In the context of epidemics, my work is two-fold. 

• Using game theory and dynamic systems theory, I have investigated the interplay between epidemic evolution and human decision-making (e.g., decisions to opt for vaccination or social distancing). Specifically, I have examined how myopic decisions, behavioural biases, and bounded rationality affect human decision-making during epidemic evolution in a mathematically rigorous manner. 
• My group has also been working on inferring network structure and predicting the future evolution of epidemics from testing data, particularly by incorporating the impact of game-theoretic decision-making (studied above) on epidemic dynamics. 

Our efforts are supported by a joint Indo-US research grant from IDEAS, Technology Innovation Hub, established at ISI Kolkata.

What prompted an electrical engineering student to focus on AI?

In my undergraduate studies, I enjoyed the subject control systems, where the goal is to control the output of a dynamic system using feedback. As I explored further, I became interested in decision theory in general, specifically optimization and game theory. I found the connection between dynamical systems and optimization algorithms exciting, specifically where optimization algorithms are viewed as dynamical systems and their convergence behaviour is established using tools from feedback control. Eventually, I studied optimization problems under uncertainty, specifically where the distribution of uncertain parameters is unknown; however, the decision-maker still needs to arrive at a robust decision or satisfies the uncertain constraints with high probability. Not only do these techniques have applications in control and other engineering systems, but these are especially relevant in AI. Machine learning (ML) since a large class of ML problems, including regression, classification, training of neural networks and generative models. My recent work focuses on the intersection of machine learning and control, especially developing efficient data-driven algorithms with rigorous statistical guarantees.

What initial difficulties did you have during the transition? How did you triumph over them?

During graduate school, I was fortunate to enrol in many excellent courses taught by highly accomplished researchers and teachers. As a result, I built a strong background in the fundamentals and, over time, began to appreciate core AI and ML research. I also had a close group of collaborators and friends who helped me along the way.  

Describe the topic of your doctoral research. What are your significant contributions to research?

My doctoral research looked at the impacts of human decision-making on the security and robustness of large-scale shared systems and networks. Traditionally, decision-making by multiple strategic individuals or entities has been modelled via game theory. In the language of game theory, numerous decision-makers choose their actions to maximize their reward functions. However, it differs from optimization problems in the sense that the reward function of an entity depends not only on its action but also on the actions chosen by other entities. Such problems arise in many social and engineered systems. For example, when a driver decides his route from one location to another, the travel time depends on his own choice of course and the route choices made by other drivers. Similarly, the likelihood of an individual becoming infected by a disease depends not only on their actions (wearing masks, following social distancing guidelines, etc.) but also on the actions chosen by other individuals it interacts with. 

The game theory had previously investigated analogous problems, but the most common assumption was that decision-makers were entirely rational, or "anticipated utility maximizers" in utility theory. This assumption made the problems analytically tractable. In contrast, human decisions are often more complex compared to such simplistic models. Moreover, several decades of past research in behavioural economics, pioneered by Nobel Laureate Daniel Kahneman, among others, have identified several systematic patterns in how human decisions deviate from predictions of expected utility theory. While experimental economists have examined these aspects, a minimal theoretical analysis of games with decision-makers exhibited the above behavioural biases. 

My doctoral research aimed to bridge the above research gap. I looked at two broad classes of game-theoretic settings. First, there are situations in which several people fight for a single resource that is prone to failure because of excessive investment. Second, I showed, in a mathematically rigorous manner, that the resource is at an increased risk of loss when decision-makers exhibit behavioural biases compared to the predictions of classical game theory models, which ignore the presence of such biases. I quantified the relative increase in the failure probability of the resource due to human decision-making models for a broad class of resource characteristics. Imposing taxes to discourage resource utilization may lead to greater resource utilization and increase its fragility, and how to appropriately design such taxes or incentives to achieve desired resource fragility. 

In the second part, I examined a class of network security problems where each node decides its level of security investment. Its risk of failure depends on its investment and the investments of other nodes it interacts with. I measured the network's robustness, studied how to construct resilient networks against multiple attack models and human decision-makers security investments, and examined how behavioural biases affect security investments.

Despite the research, what disparities do you see between Indian and Western universities?

While the focus of top Indian research universities is primarily aligned with those of leading universities in the West, there are several significant differences.

• Flexibility: Top Western universities tend to have a very flexible curriculum, particularly at the undergraduate level. The number of courses in a given semester is typically much smaller than coursework at IITs. It allows students to develop expertise in their area of interest by enrolling in subjects of their choice, learning those subjects in-depth, and devoting more time to their research projects. At the graduate level, many universities allow students to simultaneously pursue a master's in a different department (or even MBA) while pursuing their PhDs. It is common for PhD students to complete 12-15 courses during their study, which helps them build a strong foundation and later enables them to work in different fields. 
• Focus on Research Quality: During PhD research and at the time of faculty recruitment, there is a greater focus on the quality of research. Only publications at the most selective and prestigious venues serve as the benchmark. Outstanding contributions are weighed significantly more compared to the number of publications. Similarly, funding agencies in countries such as the USA often hire faculty members to serve as program managers and in other leadership roles, which enables them to identify critical research gaps and tailor their call for proposals in a focused manner. The grant review process is extensive and detailed feedback is provided to the PIs even when the proposal is rejected. Similarly, the review process is conducted on time. 
•  Networking: It is easier to remain connected with a Western university's broader research community than an Indian university. Part of it is due to the geographic distance, lack of funds supporting international travel and (sometimes) difficulty obtaining the required visa from India. Another reason is culture. Universities/ departments in the west regularly invite leading researchers in different fields to deliver invited talks and interact with students and faculty. While we do this at our institute, the extent is comparatively less.
• Empowering Young Researchers: Leading universities in the West have a culture of empowering young researchers in various ways. PhD students supported by research assistantships are often not required to serve as teaching assistants and devote more time towards research. At the same time, it is common for PhD students at IITs/NITs to do 10-15 hours of TA duty every week throughout their studies. Producing scientific equipment and workforce is much less time-consuming for the faculty members and requires much less administrative approval and paperwork. The institutional support is much more excellent. While these items may seem minor and less important, the cumulative amount of time spent on them is significantly smaller for a faculty in the West compared to a faculty member in India, which ends up being a difference maker. 

I have noticed some differences between Indian universities and their western counterparts. 

What strategies do you use to keep calm when an experiment fails, or an article is rejected?

There are always lessons to be learned with every failed experiment or article rejection. A thorough review process characterizes top journals and conferences, and often reviewers share excellent constructive feedback on strengthening the paper. I often incorporate them and enhance the contributions in my follow-up submission. Similarly, we analyze failed experiments to identify the reasons behind their failure, which provide valuable insights into the applicability of the specific methodology to the problem at hand. Either way, such outcomes improve our understanding of the problem and inspire us to develop new and improved algorithms. 

At the heart of it, research is a journey where we ask questions and try to find answers. While there are milestones in this journey regarding publications or patents, the journey should be something other than milestone-driven. 

What advice would you provide to someone who wants to work in artificial intelligence research? What should they focus on to advance?

My advice is threefold. 

• The foundations of artificial intelligence and machine learning lie rooted in linear algebra, optimization and probability. A solid background in these three subjects is necessary to understand and appreciate state-of-the-art research in AI and ML that appear in top conferences (such as AAAI, NeurIPS, and ICML) and journals (e.g., JMLR). 
• A plethora of methodologies has emerged from past research in AI and ML. However, different techniques have their strengths and limitations, including the requirement of training data and computing resources. Understanding the problem at hand and choosing the most appropriate methods for that specific class of problems is essential. The approach should be problem-centric rather than methodology centric. 
• My style is to solve a minor instance or particular case of the problem first before attempting to solve the complete problem. If a procedure doesn't work well for a minor problem instance, it probably won't work well for the larger problem instance.

What scholarly articles and publications have had the most impact on your life?

The following articles and books on stochastic optimization and viewing optimization algorithms from a dynamical systems viewpoint have influenced my recent works. 

• Bottou L, Curtis FE, Nocedal J. Optimization methods for large-scale machine learning. SIAM Review. 2018;60(2):223-311.
• Mohajerin Esfahani P, Kuhn D. Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations. Mathematical Programming. 2018 Sep;171(1):115-66.
• Powell WB. A unified framework for stochastic optimization. European Journal of Operational Research. 2019 Jun 16;275(3):795-821.

[An expanded version of this article recently appeared in the form of a book titled "Reinforcement Learning and Stochastic Optimization: A Unified Framework for Sequential Decisions " by Prof. Warrren B Powell.]

• Recht B. A tour of reinforcement learning: The view from continuous control. Annual Review of Control, Robotics, and Autonomous Systems. 2019 May 3;2:253-79. 

[This work highlights the limitation of Reinforcement Learning in continuous control problems along with an exposition of alternative approaches.]

• Lessard L, Recht B, Packard A. Analysis and design of optimization algorithms via integral quadratic constraints. SIAM Journal on Optimization. 2016;26(1):57-95.
• Meyn S. Control Systems and Reinforcement Learning. Cambridge University Press; 2022.

I have also learnt a lot from the following books and articles and used them in my research and teaching.  

• Hardt M, Recht B. Patterns, predictions, and actions: Foundations of machine learning. Princeton University Press; 2022.
• Shalev-Shwartz S, Ben-David S., Understanding machine learning: From theory to algorithms. Cambridge university press; 2014. 
• Camacho EF, Bordons C., Model predictive control: Classical, robust and stochastic. Springer International Publishing; 2016.
• Roughgarden T., Twenty lectures on algorithmic game theory. Cambridge University Press; 2016.
• Kahneman D., Prospect theory: An analysis of decisions under risk. Econometrica. 1979;47:278.
• Bullo F. Lectures on network systems; 2019.

Finally, the lifelong contributions of Prof. Vivek Borkar and Prof. M. Vidyasagar at the intersection of learning and control have significantly influenced me.