Elements of a Temporal Workforce Structure Model

Hello everybody, welcome from Sydney, Australia. We believe we're California without the guns and so a very nice place to come and sun and surf so try and find a holiday here sometime. We're very excited to present today and the the topic of our presentation is a temporal workforce model. Dave will do the presentation. I'll introduce Dave in a minute. What we've introduced is a branchial structure in time and state space. We're very excited about was a three month project to write the algorithms and we think we've done that and solved some very big customer issues. I'll turn over to Dave now to tell you all about it. Thanks Dave.

Thanks very much Stuart for your introduction. I presume you did for us folks. As Stuart probably mentioned, we're down under here in Sydney and the time here is actually 5am on Saturday morning. If you see some bleary eyes down the camera, you know the reason why. I'll launch straight into it things. We're a little shorter of time as you can see. Stuart and I work for Atura. We actually work in the data and analytics area and what we're pleased to talk to you today about is something we've been working on around developing an organizational workforce posture model. And you might say well what on earth is that? Some of you might actually already know, but what is it? Well, I mean, bear with me, we'll talk about it as we go through but in essence what it's about is an organization having the right workforce in the right place at the right time. Over time, as time progresses, this changes. Obviously it's a dynamic situation. So certainly in Australia organisations have been finding it difficult to attract and employ sufficiently skilled staff and I'm fairly certain that's the case more globally as well. Staff who are then able to be trained over a period of time to fulfill specialist or even just sort of some standard type roles are proving difficult to source and retain in some cases in highly skilled areas. In operational type industries the training and progression involved can be actually very lengthy with very complicated pathways for trainees and for training and career progression and many possible distribution or decision or turning points that occur along the way. And these are also opportunities for management intervention of course in the way they shape and mould their workforce posture. So in this environment and even more generally, the ability for organisations to be able to have some sort of predictability and level of control over their staffing profiles is very, very important, both from an ongoing operational perspective but also from a cost perspective. So if I move to the next slide here we won't go through this in detail, but the key points here are that we've developed this model using Wolfram, obviously, and the key aspect is that what this model delivers is the benefits of in fact a closed loop control over workforce posture. And the key about Wolfram here is that this relies on a reasonably complicated symbolic computation which uses convolution and multi stage optimization. As I said. What's the purpose of the model? Well, we've covered that very briefly, but many organisations, particularly across defence and similar high technology and infrastructure industries, can be construction, mining, transport, health, banking and finance. Really anything where you've got any kinds of complex training procedures in place, especially for specialised type roles. These kinds of companies have operational and capital development plans that extend decades out into the future. And as we've said, the whole key about a workforce posture, a desired posture, is the availability of sufficient personnel in the right posts, in the right roles and locations, with the right experience, at the right promotional level, at the right times. And this is what we refer to as workforce posture. So planning for this posture to tonament, if you like, which is the matching of staff supply to demand, has become an issue for organisations and particularly corporate governance. And so the model we'd like to present here, or give you an overview at least of changes the current system of management for workforce to tonament or posture control, from one of sort of an informed guesses or even an ad hoc approach based on SMEs and gut feel, to a systematic approach of the achievement of planned temporal workforce posture. So traditionally, you know, certainly the reviews that we conducted in our own experience with our clients is that temporal workforce models have been proven very difficult to build and manage, if intractable, using standard tools like Microsoft Excel and even discrete modeling packages. Because some of these things just become way too complex. What to include all of the crossing pathways and the different timings involved and the injection of your cohorts as your trainees at the top and your personnel intakes. So there's other reasons why it can be very, very difficult, including the ones that I've listed here, not least of which is sor of the natural exit from training and operational engagement profiles as people move through the profusion of all the pathways. The historical structure that exists in any organisation's workforce posture embedded within the system, which is a function of decisions and exigencies that have occurred in the past, and the material impairment due to the attrition of trained and experienced staff, which has to be very carefully managed with employment contracts and ongoing incentives. It's an Important point to note that of course a lot of the times when you've got a mature workforce structure and posture, the more senior people in your workforce are in fact those who feed back into the loop and are involved heavily in the training of either new or intermediate personnel as they pass through the system. So a higher attrition of those more senior people can have a double whammy impact. And they're obviously in a larger model, millions to billions of variables and temporal outcomes that can exist in terms of demand and supply. The taxonomy we spoke about, the demand side of workforce posture is complicated by factors that include operational tempo and dynamic contingency requirements. Things that occur over time due to contingencies, for example. And the workforce posture model we describe here does not really address the demand side, it addresses the supply side. And that supply side is complicated by six main factors and constraints. There's the long duration of individual training and operational engagement profiles, which can be a decade or even more in the case of our client. There can be many paths that these profiles can take and there's actually a non intuitive dynamics of all the sequenced end to end stock and flow effects that you, that you see in a complicated model with pathways like that. And it also includes a lot of, includes a lot of inertia and decision latency because of the time that it takes and the, and the pathways involved in such training. There's also training capacity constraints, constraints within the real system itself. And there's the attrition of trainees, experienced staff and strainers. As we spoke about, you also want to minimise the opportunity loss of under supply and the cost of pooling or oversupply of your workforce in any areas. So we spoke at the top about the fact that this model allows you to move from an ad hoc type approach into an active management approach. And so we know that because of all those effects on the supply side decisions and management decisions around employment and training need to be made years in advance of organisational posture demand to allow those people to flow through the system and reach areas of seniority that are required. And there's also, as we spoke about, the changing baseline and embedded historical structure of supply due to those sunk decisions in the past and the exigencies that may have existed. And there's an evolving demand profile over time, it doesn't just stay static. The need to review and update decisions regularly over time exists. And that of course implies an active management approach is required to the production of your trained personnel. And you have to keep an end to end perspective since the decisions that decisions you made might not be fully understood by what tends to be siloed training entities within larger corporations and larger organisations. And it's always been the case that companies reach back into universities and tertiary training institutions in order to provide that they have incentives to ensure sufficient graduates are available in the desired disciplines that they require. So if we move to the model itself, if I look at an overview of the way that the model works, we can see that down the left hand side we've got a bit of an indicator of a flow, although a lot of these things happen simultaneously. But the key element that underpins the entire model is a detailed graph of all the training pathways that exist and the career progression pathways that exist for a workforce. And I'll talk a little bit about that in the next couple of slides. But out of that graph you develop Markov pathways. So this is actually a temporal Markov. It's temporal and state Markov. It's not just a pure state based Markov process here, it's a Markov temporal and state process. That comes out of the fact that you've got a graph that describes the timings as well as the. As the states that all of the trainees pass through as they go through the system. We've then got convolution and wave valves. And this is a very, in fact, almost the most important aspect of the algorithm. We'll talk about those more shortly after we've done a convolution and a set of wave valves. The valves being points of distribution where people change direction at a point in the training continuum. This convolution and wave valves, the valves allow management input to change the values of that distribution as they pass through. But we need to do a symbolic convolution, including the way that the valves change over time. And once we've done that, and we've got a symbolic representation of how the stock changes in the graph, we then need to do a couple of optimizations. The first is involved in ensuring that you prime the pump, if you like. You need to build up that historical vintage going backwards in time in order to make sure that your stock levels match. What you know is the stock state at your time equals zero point when you're initializing the model. So that's what the pump, the pump priming optimization does. And then we move on to the posture optimization, which then says, okay, I now have a desired workforce posture that I want to see put in place over time. And again, it's not necessarily static, it's something you can Specify over time. And then we run a posture optimization to enable you to change both the intakes that occur at the top of your training graph, that the cohorts of trainees, new trainees that are coming in, and the valve settings as a function of time via the wave valve approaches. That then enables you to, as closely as possible, achieve that desired workforce posture over time. And then of course, as an output, we generate a series of time series plots to enable you to do that. But it's a non trivial process and there's a lot of simultaneous things that we need to consider and compute, not least of which is the many thousands to in fact billions of variables that are involved. Thousands of career transitions, hundreds of valves, dozens of constraints involved that we need to model, including inflow capacities and some things where you might have to restrict the flows that come through due to the talent that's available. There'll only be a certain proportion of people that have the skills and knowledge and prior training required to progress into more complicated and more demanding roles, et cetera, et cetera. So there's a lot of variables and a lot of simultaneously computational elements we need to consider. There's. So if we just move on to a look at a very basic illustrative principle of the graph itself, what we're looking at here, and this can be extrapolated to much larger graphs, which I'll also talk about. But the principle here is if you start with recruitment, you're recruiting some new trainees here, and just for the sake of it, we're going to generate sort of two pathways here. You've got nominally a pathway A, which takes a recruit through their initial training to a junior position and then through further training to a senior position. That's pathway A. You've also got pathway B. And pathway B allows people to go to another junior position down pathway B. But then if they so decide, they can then transition through extra training to do a conversion from B to A so that they also can end up in a senior position A after they've done this training at both these points here from position B and senior position A, and in fact also from junior position A, you can have attrition. Attrition can occur at certain points throughout this progression. But the main takeaway here is at this point here, people are arriving here at this particular position or state within our graph at different points in time. And that's also a function of the very different points in time that you inject your recruitment. So you're injecting your cohorts of Trainees at different points in time, and as they pass through, they're reaching points in the graph at different points in time as well, depending on the pathways that they've taken. And of course then a senior position can move on to further promotion to higher senior positions. So, as I said, the key things to note here are the various different pathways. There are lots of different interacting pathways. When we extrapolate this to a larger graph, there's specialist and standard roles. And we see this in this next slide here. This is a bit of an example, even though, again, this is nowhere near as complicated as a real world type situation. But there are pathways that take you down different pathways for perhaps specialist roles and standard roles. There are valves or distribution points. Every point here you see where there's a distribution that's required for trainees to be split to get in different paths. That's what we refer to as a valve and that's a distribution point. So you can see all these distribution points. And there are some points where there are two, there's a bifurcation, there's other points where there are more than two pathways that can be chosen. And so those valves are distribution points that play an extremely important role in the graph. And then of course, there's attrition that can occur from different points during the continuum itself. So getting down into a little bit more of the detail of what's going on here, I spoke about the importance of convolution and wave valves. Well, if we just look at the convolution here, I won't dwell on this because I'm sure people are very familiar with the principle of convolution, where you basically superimposed, sort of in an integral manner, one function over another. And in this case, what we're doing is we're superimposing the, the training pathway durations and things that are inherent in the graph structure itself with the incoming cohorts. And so you have the Dirac deltas and the unit step functions which describe those two elements. And the point, the key point being down here is that cohorts are coming in at different defined impulse points. This is entirely illustrative, by the way, these graphs, these are just illustrations of how the principle works. But you have impulses at which point you get cohorts coming in of different magnitudes. And when you run those through the graph, you can generate via this symbolic convolution and then an evaluation at points in time. And for every point in the graph, you can get a stock profile out of that and the stock profile, of course, is what you need in order to understand the behavior of a system and to understand whether or not you're reaching that desired target posture as time progresses. Again, a key point to note here is that we do this for every vertex in the graph, so that we've got control of visibility at least over every vertex, every position, every role that occurs in it, throughout our training continuum. So let me just briefly return to the wave valve concept as well. Wave valves actually allow a changing profile of all the valve settings in the graph to be implemented over time. So what this does is it allows that human intervention into the system response management decisions can be made in order to achieve that optimal outcome for an overall target workforce posture. So what the wave valves do is they actually modify that Markov structure of the graph over time. And it's actually not trivial. If you think about a given cohort entering at a certain time, then as that cohort passes through the graph, the various valve settings that experienced in the past as those valves changed over time according to a valve phasing, then as that cohort passes through the graph, the various valve settings it experienced in the past have to be carried with it, whilst also accounting for the changing valve settings for all the future possible pathways that those cohorts might experience as you go on into the future. And that has to be done in a simultaneous way. So once you extrapolate that out over a larger graph for all cohorts, over the entire timeline of what you're interested in, that becomes an asynchronous management of valves and convolutions that has to be symbolically computed and then evaluated via the optimizations we spoke about. And so what we do in terms of outputs is basically generate a lot of time series graphs. And we haven't really generated a lot here. We just wanted to show you the principle of it. What we want the takeaway to be out of this presentation is the idea that we're achieving. We're giving management the ability to achieve an optimized correction, the ability to achieve a desired workforce posture through an optimization that allows them to get as closely as possible to that. And so this management input occurs because the model tells you the optimal settings for all the valves involved in the graphs, all those distribution points over time, as well as the optimal setting for the cohort intakes that are occurring at the top. So you can almost think about this as a marble model too, where you're dropping marbles in the top and they're being distributed throughout the whole workforce structure. Over time as a function of the variables that we've spoken about. And so giving management the ability to actually see the desired optimal valve settings for all the valves across the graph, as well as the optimal intake profile that they need to try and aim for in terms of their recruiting cohort intake structure gives them the ability to see to what degree they are able to achieve their desired posture. And again here you've got optimal valves over time, you've got optimal intakes over time, but you've got the stocks that you can see over time. So you're looking at every single controlled vertex in the model, all the positions that you wish to control to achieve your posture. And you can look not only at each individual position over time and how it's going to perform based on its target number, but you can also aggregate all of those positions, whether it be by organizational element, a branch, a division or whatever, but also as a grand total over time to see what your entire workforce stock is looking like as a when compared to the target posture that you're desiring. So you can drill down in detail to look at either from a single position all the way to the, to the top level view. And that can be very powerful for management. So that leads us directly to the utility of the model and really as we've spoken about, it's the ability for management to receive timely and valuable information for decisions that have lead times of sometimes years. And the model allows you to answer these critical questions. Well, first of all, what is the confidence that we have in our workforce production system to meet our desired demand and tempo requirements? And after that there are a lot of flow in questions as well and associated questions which is things like how and when will remediation, growth and sustainment targets be achieved? What's the maximum growth potential whilst maintaining a posture? What decisions are urgent and what's the risk implications of delaying those decisions? And overarching all of that you've got what are the costs of the various strategies in order to achieve the target outcomes that the model is telling us we need to implement? And also what are the critical bottleneck capacity constraints and attrition points that we need to address? And so look, I think that's our overview of the model. I think I'll just, I'll throw to Stuart because I believe we're a little short on time. I'll throw to Stuart for some final comments and then we can may take some questions after that.

Thanks Dave. Yeah, this is a pretty exciting development. The the once developed the kernel of the graph and the convolution that then exists in symbolic form is, is like a Holy Grail and many different strategies can be tested against it. And it's a very quick process to do that. Took us three months to build the, the Holy Grail here and we're very, very proud of it. It found immediate effective use in the client and is being expanded. Its use has been expanded greatly into many areas. So we think this proves the value of the algorithm, the value of the branchial approach, the value of the impulse response of the system, the impulses being the cohorts that arrive. And so we have a completely dynamic system in continuous that allows some incredibly good insights and urgent action has been taken on the basis of our findings. Already so very excited about that and the future of what we have here. And this forms one of the pearls on our necklace of pearls. And so we have confidence analytics, we have prioritization now we have workforce optimization and we're moving on to capacity utilization. So quite a nice, quite a nice set of things that go together. One, one informing the other, this model informing whether the people side of the business can match the mechanical, the operation side of the business, whether that be mines or whether that be rolling out banking functions or something like that. So very, very proud of what we've achieved here and great presentation. Thanks, Dave. We're happy to take any questions. Thank you.

We've got a question here from Chase. He says how fine grained is the model system? For example, what's the smallest time slice for which you recompute? And what's the largest count of concurrent constraints that were applied within the model? Yeah, great questions, Chase. At the moment, the model runs on a time slice of quarters, but that in fact could be, could be reduced. I don't think you'd really want to go much smaller than that in a real world application, to be honest, but you could definitely implement it by weeks if you really wanted to go down to that.

Could I help you a little bit more there? We actually compute on a daily basis. Well, we could compute even less, but Dave's referring to the frequency of the cohorts coming in. So we have quarterly intakes with this client, but we then treat the whole of the system on a daily basis.

Apologies, misinterpreted the question. Yep, absolutely. And as far as the constraints are concerned, oh, gosh. I mean, we've got quite a few concurrent constraints because really there are constraints currently that exists mathematically within the graph at almost all vertices, certainly at any of the distribution vertices, verdicts, points and we also are adding business related constraints. We've probably got. Help me again, Stuart. I can't, I can't put a number on it actually, Chase, but it's, you know, let's say of the order of 20 other sort of business related type constraints that are put into the system. But again, there's probably really no, there's no real limit on that. The only drama with that would be the principle of how much you really want to try and constrain an optimization. So as you know, the more you try and apply more and more and more constraints to an overarching optimization, that can become a little trickier to generate results.

If I could add to that, Dave, again, we prefer to run the models in as much unconstrained mode as we can. There's absolutely essential constraints, as Dave said. There might be constraints on the number of people that have the talents to do the highly specialized roles. I mean, maybe, you know, you're looking for people that are amazing game players, for instance, for particular roles and you only want the cream of the cream of game players. And so we have constraints, but we prefer to minimize those as much as possible because models perform much better as you can, as you can minimize the number of constraints and you don't want constraints in one area of the model from affecting the posture optimizations in some other area of the model. So upstream constraints are very impactful across things. So we tend to run without constraints as much as possible and then to pull back and see the effect of putting the new constraints on. Yeah, so it's a, it's a, there's a, there's a very much an art in this. It's as much an art as a, as a science. But certainly the, you know, the holy grail we have here provides a, you know, puts all of the science in a black box that's quite small and allows you to get on with the art of the system. Thanks, Dave.

Thanks, Stuart. I don't see any other questions there on the chat. Are there any other questions from anyone? No worries, Chase. Thank you.

Okay, there's one more question now.

It was just, just Chase responding. Yes, yes, fantastic. All right, it looks like we're, we've got no more questions. So incredibly, I think we're actually a couple of minutes ahead of time. So thank you very much for attending and appreciate you coming to our talk. And I think you're almost at the end of the conference. We're almost at the end of the conference per se. But enjoy the last few present and sessions. If they're there.

Thanks, everybody. Thanks for now.