Dr. Yash Puranik: Working at Aimpoint Digital / Research on BARON

Yash graduated in 2016 from a PhD program in Chemical Engineering from Carnegie Mellon University where he was advised by Dr. Nick Sahinidis . He now works as a Senior Data Scientist as Aimpoint Digital.

We discuss Yash's experience as a PhD student, some of his research on BARON, his transition to a full-time job, and his current role at Aimpoint. Yash also has some great life advice for PhD students.

Check it out!

Transcript

0:00

Hi there, and thanks for tuning in to optimizing you the podcast where we talk about all things optimization. Interested in studying optimization, interested in applications of optimization. Interested in hearing from current PhD students, faculty, and practitioners about what they're up to date with optimization. That listen on my friend. Hey, welcome back to optimizing you. This is Anthony, and today I will be interviewing Josh paratenic. I need to say his name correctly. He is a senior data scientist at a point digital, and I'm really excited to talk to you ash, because he went to CMU, graduated in 2016 with his PhD in chemical engineering and has worked on the state of the art optimization solver baron at the optimization firm. And I think that's super cool. And we'll talk about that. We'll talk about him getting his PhD and his current job and how he's enjoying it. He's also the president of the Pittsburgh IT professional informs chapter. So it's nice to have a fellow in forms are here. Welcome. How's it going? Great. Thanks Anthony for inviting me and I'm excited to be here in chat with you. Hopefully, this is useful for the listeners. You see how it goes. Definitely all the PhD students out there listening and others. Anyone really, I feel like a lot of people have gotten good insights from the podcast, even if they're not in the optimization field. So tell me a little bit about yourself. So you did go where I got my PhD, which was Carnegie Mellon. That was that was very helpful. Yes, I did go work at the optimization for him for a couple of fields. I worked at Rockwell Automation before that. I'm currently I work for a consulting company called the important digital, where we're helping clients, or specifically I'm helping clients with any optimization challenges that they might be facing. In addition, generally, I like reading. I loved playing racquetball. It's great. I do keep coming back to CMU often to play racquetball. And I have two cats and in typical fashion, they're named perigee. Perigee because it's a cat and gigabyte. Row 1 gb. Well, as you play racquetball some time do oh, I didn't know you played. Yeah, I play a little bit. We usually play Mondays and Thursdays at seven if you want to show up. Okay. Yeah, I'm not very good, but I can play. That's good. I don't know if you need to edit it out in case there are any stockholders who may try to find us. We can play anyone who wants to. Um, so, yeah, you got your PhD in chemical engineering. And I think most of the people I've spoken with steady like operations research or industrial engineering. So were you in the process systems engineering part of chemical engineering or is that like the overlap with optimization or how did you get involved with optimization stuff? No, that's exactly right. I was involved in the process systems engineering piece of chemical engineering. And that is a question I often have to face, which is, you know, it doesn't quite jive the fact that I'm doing optimization work, but I'm also a chemical engineer by training. And I'm in the short explanation for that is that there are optimization problems in every field. Everybody needs to solve optimization in some form or the other. As far as chemical engineers are concerned, traditionally we've had to solve some very challenging problems when it comes to say, designing chemical processes or trying to solve, say, blending and pooling type of problems when you think of oil and gas. And we just needed them sold, so we got into it. And now a lot of good optimization work comes from the chemical engineering field as you know, as well as it does from other fields like computer science or even business schools. Yeah. And do you guys encounter more like non-linear optimization or like I see, you worked on barren. So I guess can you tell me about like some of the parts of the solver that you were working on with that project? Yeah. Yeah. Yes. And I think going going back to earlier comment about things being more non-linear. Yes, Traditionally when chemical engineers have had to deal with stuff, a lot of the processes we deal with a non-linear. So if you think of one of the most basic chemical engineering unit operations that we have to work with. It's a heat exchanger. We're trying to keep things up as well as cool things up from time to time. The dynamics for heat exchanger or explained by this function called L MTD, which is logarithmic mean temperature difference. But you introduce log Meyer in the non-linear domain. Seems true for blending and pooling type of problems. You're probably familiar that you end up dealing with bi-linear. There. Again, you're dealing with non-linear problems? Yeah, in terms of specifically what I worked on during my time at Carnegie Mellon, especially working with bad. And there are a couple of different things that I worked on. And probably start with the one that's most practical that I found was practical. Actually, we're working at the time with a company called air liquid, which I think later. Yeah, Every kid, which is a French company that means mixed liquid air as, as the name sounds and French. But they're in the business of liquefying it and separating it into oxygen and nitrogen and then supplying those gases to it is consumers. We were dealing with real time optimization project for them where they had a model that they had built. It was challenging to solve, and they were trying to figure it out if we could help them solve it faster and more efficiently. Turns out that the model we were facing quite a lot of infeasibility issues. A lot of the parameter values get gave rise to infeasible models. And it was really hard to figure out why things were infeasible. I mean, if you end up writing a first principles model, maybe you can look at me like, oh yeah, you probably got the sign wrong there. It should have been a plus sign, it's minus. Maybe you could do that kind of analysis. But the models we were working with were based on regression. They were database models. So now it's really hard to go look back at the coefficient and say, Oh yeah, this coefficient is wrong because they don't necessarily have, you know, there's not as much physical intuition associated with them. And in fact, even looking at first principles models is not always straightforward because as things get larger, systems are complicated. It's not easy to dealing within feasibilities by hand. So I tried to do some of the ad hoc analysis that all of us are familiar with. Drop some constraints, see what happens, maybe change the right-hand side. They were just play around with things until you figure something out. And eventually I got sufficiently frustrated with it that I thought I'm I don't want to do it by hand anymore. There has to be a better way to do this. And that actually led me down this path of trying to figure out what identify what are called IIS is. So signs for irreducible inconsistent set. It's known in other fields but different things. It's also sometimes called reducible in feasible set and minimal and feasible subsystem. What have you? There are many names to it, but the idea is to try and find the smallest possible subset you can find. Well, the idea is to try to find the smallest possible subset, but even a subset of constraints such that together they're all in feasible. But if you drop any one of them, the system becomes feasible. What does that tell you? Well, that tells you that at least within this subset, something has to change. You can not have a feasible model and tell something in this subsystem changes. The other advantages, at least in practice, even if you're dealing with really large models. Typically you can, your ISS tend to be of small cardinality. So you could potentially have a problem with millions of variables and constraints, but their eyes could be maybe five variables and five constraints. Well, that makes it more efficient to find the needle in the haystack, especially if you've just made a typo. If we just gotta coefficient wrong, it becomes faster to go identify what needs to change. So that's, that is what we came up with the heuristic to speed up by isolation and then implemented it. And Biden. I think what five different algorithms we implemented in Boston at the time. That was one big part of my PhD that thought, I think hopefully it's practically useful for everyone. So if any of you are facing and feasible mortals, please go look up how to find the S's. That's such a cool project. And i've, I've run into issues with my models where I'm like, wow, why is this in feasible when I plug in wrong? And that would have been very helpful to have. I was doing a like umpire scheduling problem for minor league baseball. If the league manager wanted certain constraints on like umpire crews couldn't see the same teams. So many times in a row. Sometimes they wanted to many things. And then the problem is in feasible because you just can't meet all the requests. And I kind of had to turn things on and off to figure out what are the issues and what can we change to make it work. So that would've been nice to have. Yeah, I think you raise a good point. You are in feasibility is not always due to errors in the model. Sometimes you're just trying to squeeze too much out of your system and that's smart possible. So an ice could still be helpful in this. In this condition where maybe you can go back and tell them that, Hey, you want X, Y, and Z constraints. All of this is too much, we cannot satisfy all of them. What's most important to you?$100 that, and we can turn off what's least important for you and try solving it again. Exactly what you're definitely a consultant. Getting, they're getting there. It's been, it's been a small journey transitioning from research mindset or consulting mindset, but I'll wait until you get to it. I'm still foreshadowing. Yeah. No. Tell me about that. I'm curious. Like even from your PhD, what kind of jobs that you apply to at the end where you dead set on industry. Did you also considering academia? Yeah. How did you, how did you get into the field you're in? That's a good question. At least when I was finishing up my PhD, I knew I did not want to get into academia. And there were a couple of different reasons for it. But I think the most important was just unlike solving interesting problems. But I'm not attached to any one of them. There's, there aren't a set of a few problems that I really cared about deeply, that I'm passionate about solving. Other than the fact that they need to be some losers related to optimization data science whatnot. I think that's an important factor, at least in my opinion, because if you were to get into academia, you're probably going to work on a series of niche problems that depending on your field, very few people in the world care about. So if you, if you want to make a career out of it and convince other people to find your research for it. You really need to be passionate about the few problems are interested in. And there were no problems in my mind like that that would keep me up at night. So it's made more sense to go into industry. In terms of home, I find how I found my first job. It was it was I did an internship with Rockwell Automation in the summer of my fourth tilde and that just kinda translate it into a job or food. So it didn't have to look too far. You know, once I got the offer at art here, this seems like a good opportunity. I know the people, the nice. I'm being paid way more than I was as a PhD student. So let's give it a shot. Let's see how it works. That's, that's kinda what took me to Rockwell Automation. That sounds really nice that you had the job offer right after that internship. I know a lot of fifth years who were like, I got to find a job and they go through all the interviews and stuff. So it's nice that you had a streamlined process there. Yeah, It was lucky how that worked out and I totally know it's not straightforward, especially finding a job after PhD. I'm sure all of you have heard this before that the importance of networks and building connections. And I think in my case, that was definitely a factor at play. Show it. I had got to do an internship at the place I was working on, but the internship offer I received was true through a networking effort, if you will, where somebody in the group. So one of my presentations about isis and found it interesting and was curious about how it would apply to one of the problems they were working on. I guess, maybe stepping back a little word. The point I was trying to make here was It's important to keep that in mind, especially when you go to conferences, e.g. the first couple of conferences I went to, I thought, hey, this is fantastic. I get to do a few talks, listen to a few talks, but I get to travel on my advisor's advisor's find and it's kind of like a vacation in some sense. And this is so awesome. There's no harm in having fun at conferences. But it's also important to remember that the people who are attending with potentially going to be your pills or collaborators, they're potentially going to be our bosses when we offered you a job. So it's important to keep that in mind and try to make connections whenever you get the chance. Yeah, I like that point about networking. And I especially like that about in forums. I mean, that was the first conference that I attended was in forums last What was it? October. And yeah, it was great to meet so many people. And then I went to another conference in forums Computing Conference in January. And I saw some of the same people. And it felt nice to start like building relationships that I know I'll continue to have for years and years because we're in this kind of smaller field and there's only so many of us working on these things. Yeah, exactly. And you know, they get to see your evolution too, because when you go in as a first-year, second-year student versus when you go to a conference as a fourth order fifth year student. Hopefully there's a tremendous good that you have, that you have progressed on. And it's nice if others can also get to witness it. This conversation is a little tricky for those of those of our audience who are probably doing their PhDs and covert diodes, which meant a lot of inputs and conferences did not happen. And networking is really hard on Zoom. Hopefully, as we come out of it, it becomes more and more possible. Yeah, definitely. I think I struggled during the COVID years not having enough human connection. So even though we did some things on Zoom, Yeah, it's good that we're back in person for the most part. Fingers crossed. That continues. Fingers crossed. Let's launching So no. Can you tell me more about what you do in your current role with aim point. And like I know you do consulting and work on optimization, but like what, what's your team look like? What kind of projects you work on, stuff like that. Yeah. So important digital is a fairly young consulting company, was founded in 2017. So turning five soon turning five and may actually happy birthday. Oh, pink here. And I'll make sure to let our founders know. And we've kind of grown from say, a couple of the original foreign to us to now a team of about 70, growing pretty rapidly, which is awesome. Another nice thing or the thing I find very interesting or important is that we will split into a couple of different teams. So there's a data analytics team or data engineering team and data science team. And through a combination of all of these teams, we had all, all aspects when it comes to analytics or data science or optimization. So what do I mean by that? E.g. let's just say you're a fairly new company in the space, or maybe you're a traditionally held company that has relied on data a lot. That's when maybe I could use some of our data engineering team skills where we can come in and figure out what kind of databases you need, build the right data pipelines, ETL pipelines and so on and so forth. Cool. Now let's just say you have all the data in place, you know, where it's coming from, you know where it's saved. But by itself is not very useful, right? Maybe you want to try and figure out just what does the data mean, get some insights out of it. That's kind of what a lot of the work that our data analytics team does, it does a lot of reporting work automation would trying to convert the data that you have into insights, into dashboards that say maybe makes sense, that I run periodically that can help you identify some things. Data by itself, as I said before, it's not useful. But now, if you can come up with a dashboard that points out like, hey, you don't have five plants. Four of these plants perform really well. One of these plants has more downtime whereas needs more maintenance and the others, what's going on? That's when maybe you can go into what our data science team does, where they'll try and help you build predictive models, machine learning models and tried to figure out like, yeah, Plant fire does seem to have a lot of downtime. Let's see if we can figure out why that's happening. Or maybe let's see if we can predict when maintenance issues might arise next, next, next. So then you can try to deal with it proactively. E.g. me, specifically, I sit in the data science team and if you think about it, Analytics is a more recent buzzword. Obviously, everybody is into it in that context where it is optimization. So set, at least the way I think of it. It's going from, say, predictive analytics to maybe a more prescriptive sense, optimization results. We would like the results of an optimization model to actually just give you straight the decision that needs to happen. E.g. like the MLB case you're talking about, you just solve a model. All the empire assignments are made. We are done, We're done, we move on. That's the end goal. Should that doesn't always happen. We know that in all cases we can have an optimization model that immediately gives the results that you can or decisions that you can implement. But at least it takes you one step closer. It takes you one step closer from, say, the data, the data engineering barrels or from the descriptive analysis that maybe the data analytics team does to maybe some predictive modelling that the data science team is doing. Optimization is taking you to the next step, hopefully getting you some more prescriptive answers. So that's kind of the dataset. The optimization fits into the data science perspective. In terms of the kinds of problems I work on. All projects I work on, really depends on the client's important. Digital has traditionally had a lot of work in fields such as logistics, retail, CPG manufacturing, biopharma, e.g. depending on the client needs or projects or engagements can be as small as one week. You just need a very if the model that you need, the results that you need are not super complicated, if we can do it in a week. Fantastic. Why not? And some cases the engagements are larger. So there's a project I'm currently working on for production scheduling for. Pharmaceutical company and that's a longer engagement. No two projects are the same. That said there are some parallels and that at least I'm building and solving optimization models. The teams can vary for smaller engagements, that can just be a team of one where I love all my coworkers. Or in, in larger teams. It could, could be maybe two or three or four of us where we are trying to work on different aspects of the optimization model together. And that's one thing that's quite interesting, working in consulting or just working with team members. So I'm curious. And three, when you said you were working on the MLB empire assignment problem, how many of you were working on that model? Just me. Just you and I loved all my coworkers. That's good to hear. We had we had other students working on similar problems with other leagues, but one student tackled one league. And then we had maybe five or six legs. So we had like, yeah, some people do an umpire scheduling, some people doing the actual game scheduling. So we helped each other out, but you were responsible for your own deliverable. Now, think about it. If you had to work with somebody else from being partially linked problem, how will you have how would you have done that? How would you have split your work? Well, we would need GitHub. Yeah. Sure, sure. And I guess I it's tough because on one hand it would have been nice to sit down next to someone else. Like I had this friend James stroud who was like really bright or this guy Neil. And it would've been good to sit down and do the coding together in the modeling on the whiteboard together and do all that. But if that's wasting time and we wanted to break it up, I would have been more difficult because I would have had to write some things codes and things, have a version of the code running and then explain that to kneel. And then he would have to understand that and asking you questions and then iterate from there. So I would prefer doing it together simultaneously. I think you touched on some of the challenges. I shouldn't call them travel is just with just some considerations that we face in practice. And that is something that students typically don't face during their PhDs. But assuming they go to industry and most likely you're never going to work by yourself. You're going to work in teams when you're trying to solve a problem, that becomes something you have to learn how to deal with. So e.g. one thing you mentioned, yes, using version control to share code, whether it's GitHub or GitLab or whatever you like, doesn't matter, but I agree that's definitely something important. One thing we tried to do is okay, if there are three of us on the team. If three of us simultaneously discuss our model, yes, we absolutely need to do that. We need to make sure that all of our constraints makes sense. All of our variables make sense that nothing is more complicated than it needs to be. Are there any redundancies and the morals of the writing? We definitely need to brainstorm that and discuss that together. But once a model is finalized in some sense, or at least a preliminary model, we have an understanding of what we want to try. It doesn't always make sense for three people to be editing the same piece of code. You're probably wasting time. Now the challenges trying to make sure that, you know, how do we split an optimization model in meaningful parts that each one of us can work independently and then merge it together. That makes, still makes logical sense. That gets very interesting. Another interesting piece is that if you, especially if you're working with a client, depending on their background, some clients have optimization backgrounds and they just need help to speed things up on their end. Others don't necessarily have optimisation background, so it's not clear to them what the process entails. Now if you tell them that, oh yeah, we have this complicated model that's going to take five weeks to fully build out. We'll build it and give you results at the end of five weeks, they're going to get very, very anxious around it. They usually want to see results in an incremental fashion as they are, as, you know, as you're developing things. So at least one of the things we tried to do for scheduling model, e.g. they simply was we tried to break the model down into meaningful parts. So again, going back to your MLB example, maybe the most basic thing you need is that every, every game has an empire assigned to it and you can't have two Empire, e.g. you can have one unpaired assigned to two games on the same day. That won't make sense. So let's just build the first boost simple model with only that constraint and see what the results look like. Maybe the next step you do is maybe write some kind of as if you had some kind of stuffing restriction where you can't have an empire work every day for two weeks straight, because I like that better word. With them too much exhausted and read through much. Yeah, exactly. So maybe somebody else can work on trying to reduce that or put that limited on how many consecutive games they could work potentially. That is, I think you mentioned I was a constraint where you cannot have an umpire just too many games with the same team? Yep. Because we have potential for bias or whatever the random adults don't even out. So maybe another team member could be working on implementing that constraint. If you're careful on how you work together, hopefully are codes can be much right away and there will be no conflicts. That's where you can split the world and conquer. But at the same time, doing things incrementally in that fashion is very important to make sure that you are making sense of things as you're developing them in B, you keep showing something for your client because, you know, it's it's important to keep them apprised of your progress. So I kinda forgotten what the original question was. To be honest. Hopefully this conversation though is going in a direction that's still useful. Yeah, I I still find it interesting. I think working with clients is definitely interesting and working on a team is interesting. And yeah, there's a lot of ways to do that. I was also thinking maybe one person could do like write a bunch of tests while the other person writes the code. And then you have that kind of independent testing versus implementation so you can double-check your work in that sense. The other one maybe like you're saying, the clients might not know optimization at all. Or maybe one person is better at speaking to the client about your model and showing them some incremental stuff. So maybe one person could be more like front-facing while the other person does more of the dev work? Yeah, I guess it depends on your team. What's your team like? When you work in some deans? Go ahead. No, no. Those are both good, excellent points, especially for us. The b is still a very small company with a small team in the data science team, e.g. so it varies from project to project. That out. I think what 14 of us in the data science team total, depending on the project, will be grouped into groups of two or three or four. So I don't think I have worked with the same person more than once as of now. I've only been at important for about ten months. That may change. I'll let you know if I have favorites then. But now I think, I think you're right. What you said about testing versus the event, splitting them and keeping a separation. So you can actually double-check your work in a good way. That's a great way to do split. I will say if you're thinking of getting into consulting, you can't get away without interfacing with clients for too long. So yes, presumably one person could be better at managing the relationship than the other, but everybody just has to get comfortable with it. I think that's just something broader. What PhD is that? I did not think of much when I was doing my PhD, but it has become more and more important even before I got into consulting, is that we need to be able to present our results to non-technical audiences. We need to be able to communicate that well. Even if even when we were working when I was working at Rockwell Automation, e.g. I. Wasn't a research group. We're just a bunch of PhDs that we're working on interesting problems. And we could talk to each other very, very freely. And it was, it was quite a lot of fun. But every once in a while, you do need to present our work to business folks. You will need to present our folks to managers who may not follow or who may not have the same background as you. But at the same time, the work you're doing potentially is practically relevant to them. It's being able to communicate to them what exactly you're working on. Why is this important? Why is this relevant to them? What does progress mean? E.g. just the fact that you had an optimization solution before and then you improve it. Five was sent in terms of some, some arbitrary objective function you came up with. Well, what does that mean to the business or CO? This is where things get start getting relief fund. Say there is a lot of uncertainty in your model and now you're doing some kind of stochastic optimization. I'm sure you're familiar that if you compare the deterministic model with a stochastic model, typically the two-state stochastic models, e.g. is going to lead you with a word subjective than the deterministic equivalent. Now how do you explain that to a manager that you know what the stochastic models, so it makes more sense, although your objective looks worse, don't worry about it. These, these, these kinds of questions become very important in practice. Especially if you want your work to generalize if you want or work to be applied. If you want to work to make a difference and something I didn't think about much during my PhD, but it's becoming more and more relevant now. So something to keep in mind as you're going through your PhDs, hear me. Yeah, that's a good point. I don't think I've developed that skill too much. Do you have like, what are some good ways that a PhD student could maybe work on that to be more prepared if they want to go into consulting or a similar field. I think the first question, the first suggestion would be just practice. Practice different ways of explaining a concept. I'm sure you have friends and other departments e.g. who are still technical but don't have the same background as you. Try out material with them. Ask what they think, how, how they can make sense of a point better. Another really important piece is to always try to get a sense of your audience before you're presenting and try and tailor your message as much as possible to the audience. Another example, if you go to the informs annual meeting, e.g. if you're trying to present our work on optimization models, you don't have to, you don't have to worry about anything you can reasonably expect that an informed audience, It's fairly familiar with basic concepts of 4D. A couple of weeks ago I was at this conference in Boston called Open Data Science Conference, as the name suggests. So Data Science Conference, so a lot of machine learning practitioners, ML engineer, Simon students for that. Then if you start talking to them like, oh yeah, we solve optimization models are rebuilt models. For them. The word model is traditionally associated with the machine-learning model, maybe a neural net, maybe an NLP model. So yeah, if you start using the word model for optimization, you've lost them already and bridging that gap becomes too hard. So keeping that in mind, trying to figure out where the similarities lie, there is no single answered. Just have to keep practicing and keep trying out new ways and hopefully we keep getting better each time. We try like that. Yeah, definitely think knowing your audience is important. Sometimes they'll go into a talk and not really have that in mind. And then you see the blank stares once you start talking about something too technical. And it's like, oh boy, I got to reel it in there. Kind of pivot. So, yeah, it's good to gauge the audience beforehand. Yeah, Another example is going to an informed audience. You can talk about branch and bound as branch inbound and chemo, that's okay. They're not fazed by it necessarily. At the Open Data Science confidence, if you use the word branch inbound, one of the things that they want immediately think of as exponential complexity and be hard, this is too hard to do. Why should you even doing it? Why are you even thinking of it? You still have to make sure to either presented in ways that make sense Or remember that traditional computer scientists may or may not be as comfortable using an algorithm that has exponential complexity. So maybe making the point that yes and the worst-case scenario, this cannot be solved, but we can still solve a lot of practical problems in reasonable amount of time that the system useful to work on. Now you have to keep that in mind, which you probably don't have to worry about when you're talking to an informed audience. So just just practicing that can be thinking about that can be helpful. Yeah, definitely. It sounds like you have a good grasp on both the ML field and the optimization field, which I feel is like a great combination. Actually, my advisor has been trying to tell me, I should learn more ML. So I have both sides of the coin there. What do you like? How did you how did you get into both like, and how do you think they play well together? Well, firstly, I just want to point out that I'm a consultant. It's my job to make it sound like and do everything. Otherwise, I want to stay in business for too long. So we'll just keep that point aside. But jokes aside. Though my background or training or research was more obviously an optimization. That's what I spent time formerly learning. Ml or at the Maxwell, your diamond has been more applied, more practice-oriented, just trying to figure out how to solve some interesting problems and ended up picking up things along the way. That's actually another nice thing about doing a PhD is that, or at least in my case, when I started my PhD, I had no idea how to do research. Just just how to do research is so interesting and different from anything that you're used to in undergrad. It's definitely not the same as doing classes and getting good grades on them. That helps being able to assimilate material delivered in the classroom is important for researchers. Pigs, you know, it takes a lot more effort than that. But stepping back from that, at least, something I learned during my PhD is just figuring out how to if I have a problem, e.g. just trying to figure out what do I need to solve it? What are the resources I may need to go to to figure out what I need, what I need to do to solve it. Or just figuring out fast what, what from the literature could be useful and more from the literature is probably not going to be useful to solve my specific problem. I think that is a very useful skill that I was able to practice during my PhD and that definitely helps node in the job. So e.g. at Rockwell, one of the projects we were working on had to do with trying to deal with text data. Obviously don't have a lot of formal natural language processing, painting. I wouldn't say the word NLP because that probably means something different for this audience. So I was able to narrow down first on some of the methods that I thought could be useful for the problem we were solving. Found this one specific book that was well-cited that had that did go over the specific methods I was interested in because able to read it over, we can then figure out a solution for it. And that just something hopefully all of you will get to, or maybe you can do it even faster than me. At the end of your PhD, just figuring out what you need to solve your problem. Yeah, that, that's kind of my exposure to ML. That's how I have been playing with both ML and optimization. Part of it also is that typically when you're working on optimization models, something it seems like things are given to you. Like, you know, e.g. again, going back to your MLB and by channeling example, you probably knew the shadow will write, you probably knew when the games were happening and who was playing, that. That information was fixed. In a lot of cases though, when we assume that our inputs to an optimization model or fixed, really fixed. If you're thinking of production scheduling and if you're thinking of like, oh yeah, this is my demand. This is how much I need to make next year. Is that the mind really fixed? Is it static? But we know how it's going to change or no, right? Or we think of prices of raw materials and things like that. Well, how many people predicted the war in Ukraine? And probably a lot of people did, but the price shows that it has sent. And where are you able to exactly know what the price of food would be today? Would you know what it would be six months from now? Probably not. So we assume that a lot of our inputs for optimization are fixed and in practice they're not. Send that context. Ml and optimization are really important to think of together. Because since you don't know, are these inputs, since you're, for a lot of these inputs, you are building models. You are doing some kind of forecasting. You are doing some kind of predictive work. That is what is feeding in interior optimization model. So thinking of the synergy, thinking of how they interact together, it's just, it's just necessary in practice. Yeah. That's that's an interesting point that there'll be cases like for me, everything was fixed up front and nothing could really change or nothing could go wrong. It's just like, yeah, very, very determined. And I guess what you said at the end there, like oftentimes the ML models produce predictions are something that become like data or like coefficients for your optimization model. And I think, I like, I wonder if people are trying to blend them together even more than so. It's not like two steps of the ML than the optimization if there's some way to bring them together more? I would think so. I think that's an interesting problem that at least some people are looking at. I can't think of any references or any literature from the front, but all of my mind, but specifically relates to this problem. Let's definitely interesting and it will also be interesting to Norfolk can even be done together. Theoretically isn't even. What does it mean to do those things together? In what contexts we can do it together? Or I mean, as a practitioner, I don't really care about ME theoretical limits as long as I have good ways to do it fast. But it's, those would be interesting problems to think about. Like the, like, well, I know mLs definitely used in optimization in terms of when you're doing like even an entire program like branch and bound. They, they're coming up with better ways to figure out what to branch on and what types of cutting planes to add and stuff like that. So I like that synergy to between the two fields. Yeah, I mean, ML is, relies on optimization right here, minimizing some, either metric, some L1 norm, some L2 norm. It is an optimization problem that you're solving to build a machine learning model. That soon as you do, definitely, they go hand in hand. The other thing that you mentioned was like learning how to do research. I think that's an interesting topic and something that I didn't expect to be difficult or something I needed to practice until I started doing it because I didn't do much research in undergrad. And then I went and worked in finance as a software developer. And then I came to start doing research just because I thought, oh, research, that sounds fun. But I didn't realize like what you're saying. It is tough to identify what, how to tackle a problem and how to do a good literature review and stuff like that. So how did you like did you progress a lot from your first year and your PhD till the end in terms of being good at research and how did you do that? I would hope I progressed because otherwise it would put into some questions about the degrees C we awarded me. I'm going to call up neck. We'll see what Nick has to say. No. It stuff and I don't think I did progress eventually. I think the first two or three years of my PhD work hard. I did not feel like I was making any progress. Did not have any papers written by the end of three years. Maybe I had one draft ready to go out, but nothing was published yet. And it can be tough. You know. Ideally, you will know that it does not make sense to compare two PhDs, even if others who joined the same year as you or even who may be working on similar problems? No, to PACs are the same. Everybody has a different PhD experience. Everybody solving a different problem. Some problems are easier than the other. Sometimes you just don't know what you don't know. You may try an approach that looks promising, but as you dig into it, it turns out that it does not work for your problem. And for whatever reason, we don't seem to appreciate null results as much in the field as we should. The fact that you tried and approach and we're able to figure out that it does not work, is still valuable information. But for now, I'm going to get the paper out of it. Such as life. That was hard, keeping in mind that no two PhDs are the same. Trying to deal with impostor syndrome, that was definitely hard for me trying to figure balanced the fact that some others, some of my peers who started with me had already had a couple of papers by the end of their third year and are still struggling to write my first. It wasn't easy. I think. What helped you what helped lead you to put into perspective is that all the work I was putting in in the first three years that eventually paid off. My last two years were a lot more productive than the first three. And that's just something to keep in mind to keep having faith in your efforts. Again, we don't know. What we don't know is all you're going to succeed. Probably not. Nobody can predict that. Just have to persevere through it. In that context, I do want to give a shout out to gaps at CMU. Caps stands for Counseling and Psychological Services. That's kind of the organization at CMU that's available for students. They're facing any challenges, any issues with their studies, with their research in PCs are hard. I'm sure anybody who's done a PhD has gone through phases where things have been very stressful and very overwhelming. And that contexts, since we have such a great resources gaps here at CMU or I'm sure better at equivalent to other universities. I think it's a good idea to exploit them and meet them. Going back to your question though about how did I say improve on how to do research, e.g. I feel like we're doing this a lot. You asked me something and I go on tangents. I appreciate so much your tangent though, because that was such a good one like yeah, like hearing other people say they have like imposter syndrome or they struggle at the beginning of their PhD. Just like knowing other people go through similar things, I think is really helpful for everyone to feel better about themselves. Like we're all in it together, we all are learning and there are good resources and stuff. So I like your tangents right now, but maybe we can talk about how you can edit it and just throw questions in so it doesn't seem like I'm going off on tangents too much. Nope. No, I think it just became bit better. I think over time it was just a matter of practice. So when I, again, for stereo, when I started reading some optimization peoples, it would take me even a week sometimes to get through a paper and to understand exactly what what it was trying to do or what problem they were trying to solve, or how it made sense. As you become more and more familiar with the material. As I became more familiar with it, things got a lot faster again. Now by the end of the, by the end of the fourth pillar, when I read, read a paper, I didn't actually had to read it entirely to know what they were doing. Sometimes I could read very quickly the, the paper, say in an hour and know exactly what they're trying to solve and maybe decide that. It's interesting, but I don't need to look at it right now. Practice does improve. You just have to keep faith in the process. It also helps if you have a great advisor and if you have a good relationship with their advisor, that's for sure. So again, a very special shout out to Nick, my advisor, who, you know, there's, there's a, there's a clear power dynamic at play. There's a clear imbalance with neoplasia relationship. Phd advisors hold a lot of power would be easy students. The good ones usually are mindful of that and are trying to work in the best interests of the students. And Nick was definitely one such advisor, in fact, had such a great time working with them. There was a part of the reason I decided to come back and work for him in the optimization for it. Some would say that my Stockholm Syndrome was too strong. I don't know. But it's important, it's important to have a good advisor. That's something that can be a little bit of micro flux sometimes. But if I finding a good relationship with your advisor or maybe not just a PhD advisor, but having a network of mentors who can help you with your research process, could be your advisors, it could be your PhD committee members, it could be just other peers, e.g. is very, very useful in that context and that can go along way. Minor point also is just being able to figure out how to put more structure in your life. So again, going in college, the structure was enforced and there are classes, timetables are fixed at the beginning of the semester, your exams are fixed and you kind of followed that PhD short. Maybe you're doing a few classes in your first year. After that, you're kind of on your own, right? You just do research and write papers when you're done with it. And that lack of structure, at least for me, initially, was quite, quite challenging. There were days when I didn't start working until until the afternoon. There were days I would stay up late in the night. There were days I did not work at all when I should have. In hindsight, that was not beneficial at all. So one of the changes that did happen during the third and fourth year of my PhD, I started having a very rigorous routine of my daily schedule of when I would work, when I would stop working, I also started doing it very rigorously that I'm not pushing work on the weekends at all because you need to take care of your work and life both together. And that is something I wish if I could go back is just trying to put some more structure and more discipline in my work habits. If I had done that sooner, I think that would have been more beneficial for me. You're preaching to the choir on that one, I'm like the most structured human ever. Like. I like to work 9-5, the rest of the students in my pod or who are kind of like in my office or like, how how are you always in here at night and they're like, did you have class this morning? I'm like, No. I just get in at nine and leave around five. And I think that structure, I kinda got this structure from working before my PhD. And I think for a lot of students, you don't, you don't learn that nine to five kind of mentality. And also maybe that isn't for everyone and you can structure your schedule in whatever way you want or if you want to take a day off or start later and earlier, you know, it's all good. But I like the point about structure for sure, right? And I mean beasties or belong, you know, you're getting older as you do them. Other considerations can come into play. You may have partners, you may get mad at some PT students have kids having structured as important to balance everything out. So I'm guessing, Anthony, since you you're very religious world working nine to five, you do find some time for hobbies or interests. What are some of those? Yes, I do some I do a bunch of things. I'm wondering Greek right now. Your advisor might be happy to hear that. Yeah. Yeah. My dad never taught me as a kid. Come on, dad. Now on Wednesday nights I take Greek language class. I also play some piano. I do CrossFit, which has been a fun new hobby that keeps me active and it's fun. I've also met randomly some CMU professors or postdocs who worked out at the same gym as me and I didn't realize they're from CMU until recently. And I had emailed one of these professors before. And then the other day I was like, I was talking to this guy. His name is George Cantor. He's in robotics, does like agriculture robotics. And he was like, Yeah, I work at CMU. I'm a professor and I was like, no way, what's your last name? And he said Cantor and I was like, I sent you an email three weeks ago. I like doing those kind of things outside of class where I get to meet people who otherwise I wouldn't run into. Yeah, and that leads us to a timely PSA. Be careful of trashing your professors and public. Never know when you might run into them. That's a great point. That's important, right? Because I think that is, there's clearly some diminishing marginal return at play with research. I don't think personally, I can do very heavy complex mathematical work for more than four or 5 h in a day. I just physically can't I can spend more time on the computer on my desk. That won't be very helpful. Instead of that kind of structure that you have signed a very specific period of time for work. And you're making sure to take care of your other interests, hobbies, finding happiness in life. That's important, that that does add up to a PhD experience. I try my best to always work out the best way, but yeah, I try. I guess Anthony didn't talk about dancing, but he did show me some cool moves when I got to the studio. So maybe we need to do a YouTube video at some point. Forecasts, not going to be very helpful for that. Luckily, they can only hear voices. I don't have to demonstrate my dance moves. Not yet. Well, this has been a great a great time talking to you and hearing about your work. And you have great personal insights and great advice to students about like work-life balance and learning how to do research. So this has been very enjoyable for me. Do you have anything else you want to add or say before we wrap things up? No, I'm glad you found this useful. I hope. I hope anybody else listening to this also finds it useful. If I'm always happy to chat with folks. If my expedience for any of my insights can help you in any way, please feel free to reach out. Probably find me on LinkedIn very easily. I don't think there are too many years per annex working at the endpoint that are in because only one. That said yes, if you did not like this episode or if you did not find anything I had to say interesting. I would appreciate no hate mail. Please. Just reach out if you found this useful. Amazing. No hate mail to me either. If you don't drive a bad guess. But definitely link up with either of us on LinkedIn or send us an email or something. If you want to chat more. It's been great having you on. Yes. Thank you very much. Thanks. Bye, everyone.