Architecture of a decision-support system
(PuneTech is honored to have Dr. Narayan Venkatasubramanyan, an Optimization Guru and one of the original pioneers in applying Optimization to Supply Chain Management, as our contributor. I had the privilege of working closely with Narayan at i2 Technologies in Dallas for nearly 10 years.
For Dr. Narayan Venkatasubramanyan’s detailed bio, please click here.
this is a follow-up to optimization: a case study. frequent references in this article to details in that article would make this one difficult to read for someone who hasn’t at least skimmed through that.
a layered view of decision-support systems
it is useful to think of a decision-support system as consisting of 4 distinct layers:
- data layer
- visibility layer
- predictive/simulation layer
- optimization layer
the job of the data layer is to capture all the data that is relevant and material to the decision at hand and to ensure that this data is correct, up-to-date, and easily accessible. in our case, this would include master/static data such as the map of the field, the operating characteristics of the helicopter, etc as well as dynamic data such as the requirements for the sortie, ambient conditions (wind, temperature), etc. this may seem rather obvious at first sight but a quick reading of the case study shows that we had to revisit the data layer several times over the course of the development of the solution.
as the name implies, the visibility layer provides visibility into the data in a form that allows a human user to exercise his/her judgment. very often, a decision-support system requires no more than just this layer built on a robust data layer. for example, we could have offered a rather weak form of decision support by automating the capture of dynamic data and presenting to the radio operator all the data (both static and dynamic), suitably filtered to incorporate only parts of the field that are relevant to that sortie. he/she would be left to chart the route of the helicopter on a piece of paper, possibly checking off requirements on the screen as they are satisfied. even though this may seem trivial, it is important to note that most decision-support systems in everyday use are rather lightweight pieces of software that present relevant data to a human user in a filtered, organized form. the human decision-maker takes it from there.
the predictive/simulation layer offers an additional layer of help to the human decision-maker. it has the intelligence to assess the decisions made (tentatively) by the user but offers no active support. for instance, a helicopter scheduling system that offers this level of support would present the radio operator with a screen on which the map of the field and the sortie’s requirements are depicted graphically. through a series of mouse-clicks, the user can decide whom to pick up, where to fly to, whether to refuel, etc. the system supports the user by automatically keeping track of the weight of the payload (passenger+fuel) and warning the user of violations, using the wind direction to compute the rate of fuel burn, warning the user of low-fuel conditions, monitoring whether crews arrive at their workplace on time, etc. in short, the user makes decisions, the system checks constraints and warns of violations, and provides a measure of goodness of the solution. few people acknowledge that much of corporate decision-making is at this level of sophistication. the widespread use of microsoft excel is clear evidence of this.
the optimization layer is the last of the layers. it wrests control from the user and actively recommends decisions. it is obvious that the effectiveness of optimization layer is vitally dependent on the data layer. what is often overlooked is that the acceptance of the optimization layer by the human decision-maker often hinges on their ability to tweak the recommendations in the predictive layer, even if only to reassure themselves that the solution is correct. often, the post-optimization adjustments are indispensable because the human decision-maker knows things that the system does not.
the art (and science) of modeling
the term “decision-support system” may seem a little archaic but i will use it here because my experience with applying optimization has been in the realm of systems that recommend decisions, not ones that execute them. there is always human intervention that takes the form of approval and overrides. generally speaking, this is a necessary step. the system is never all-knowing. as a result, its view of reality is limited, possibly flawed. these limitations and flaws are reflected in its recommendations.
this invites the question: if there are known limitations and flaws in the model, why not fix them?
this is an important question. the answer to this is not nearly as obvious as it may appear.
before we actually construct a model of reality, we must consciously draw a box around that portion of reality that we intend to include in the model. if the box is drawn too broadly, the model will be too complex to be tractable. if the box is drawn too tightly, vital elements of the model are excluded. it is rare to find a decision problem in which we find a perfect compromise, i.e., we are able to draw a box that includes all aspects of the problem without the problem becoming computationally intractable.
unfortunately, it is hard to teach the subtleties of modeling in a classroom. in an academic setting, it is hard to wrestle with the messy job of making seemingly arbitrary choices about what to leave in and what to exclude. therefore, most students of optimization enter the real world with the impression that the process of modeling is quick and easy. on the contrary, it is at this level that most battles are won or lost.
note: the term modeling is going to be unavoidably overloaded in this context. when i speak of models, students of operations research may immediately think in terms of mathematical equations. those models are still a little way down the road. at this point, i’m simply talking about the set of abstract interrelationships that characterize the behaviour of the system. some of these relationships may be too complex to be captured in a mathematical model. as a result, the mathematical model is yet another level removed from reality.
consider our stumbling-and-bumbling approach to modeling the helicopter scheduling problem. we realized that the problem we faced wasn’t quite a text-book case. our initial approach was clearly very narrow. once we drew that box, our idealized world was significantly simpler than the real world. our world was flat. our helicopter never ran out of fuel. the amount of fuel it had was never so much that it compromised its seating capacity. it didn’t care which way the wind was blowing. it didn’t care how hot it was. in short, our model was far removed from reality. we had to incorporate each of these effects, one by one, because their exclusion made the gap between reality and model so large that the decisions recommended by the model were grossly unrealistic.
it could be argued that we were just a bunch of kids who knew nothing about helicopters, so trial-and-error was the only approach to determining the shape of the box we had to draw.
not true! here’s how we could have done it differently:
if you were to examine what we did in the light of the four-layer architecture described above, you’d notice that we really only built two of the four: the data layer and the optimization layer. this is a tremendously risky approach, an approach that has often led to failure in many other contexts. it must be acknowledged that optimization experts are rarely experts in the domain that they are modeling. nevertheless, by bypassing the visibility and predictive layers, we had sealed off our model from the eyes of people who could have told us about the flaws in it.
each iteration of the solution saw us expanding the data layer on which the software was built. in addition to expanding that data layer, we had to enhance the optimization layer to incorporate the rules implicit in the new pieces of data. here are the steps we took:
- we added the fuel capacity and consumption rate of each helicopter to the data layer. and modified the search algorithm to “remember” the fuel level and find its way to a fuel stop before the chopper plunged into the arabian sea.
- we added the payload limit to the data layer. and further modified search algorithm to “remember” not to pick up too many passengers too soon after refueling or risk plunging into the sea with 12 people on board.
- we captured the wind direction in the data layer and modified the computation of the distance matrix used in the optimization layer.
- we captured the ambient temperature as well as the relationship between temperature and maximum payload in the data layer. and we further trimmed the options available to the search algorithm.
we could have continued down this path ad infinitum. at each step, our users would have “discovered” yet another constraint for us to include. back in those days, ongc used to charter several different helicopter agencies. i remember one of the radio operator telling me that some companies were sticklers for the rules while others would push things to the limit. as such, a route was feasible or not depending on whether the canadian company showed up or the italian one did! should we have incorporated that too in our model? how is one to know?
this question isn’t merely rhetorical. the incorporation of a predictive/simulation layer puts the human decision-maker in the driver’s seat. if we had had a simulation layer, we would have quickly learned the factors that were relevant and material to the decision-making process. if the system didn’t tell the radio operator which way the wind was blowing, he/she would have immediately complained because it played such a major role in their choice. if the system didn’t tell him/her whether it was the canadian or the italian company and he didn’t ask, we would know it didn’t matter. in the absence of that layer, we merrily rushed into what is technically the most challenging aspect of the solution.
implementing an optimization algorithm is no mean task. it is hugely time-consuming, but that is really the least of the problems. optimization algorithms tend to be brittle in the following sense: a slight change in the model can require a complete rewrite of the algorithm. it is but human that once one builds a complex algorithm, one tends to want the model to remain unchanged. one becomes married to that view of the world. even in the face of mounting evidence that the model is wrong, one tends to hang on. in hindsight, i would say we made a serious mistake by not architecting the system to validate the correctness of the box we had drawn before we rushed ahead to building an optimization algorithm. in other words, if we had built the solution systematically, layer by layer, many of the surprises that caused us to swing wildly between jubilation and depression would have been avoided.
other articles in this series
this article is the second in a series of short explorations related to the application of optimization. i’d like to share what i’ve learned over a career spent largely in the business of applying optimization to real-world problems. interestingly, there is a lot more to practical optimization than models and algorithms. each of the the links below leads to a piece that dwells on one particular aspect.
articles in this series:
optimization: a case study
architecture of a decision-support system (this article)
optimization and organizational readiness for change
optimization: a technical overview
Dr. Narayan Venkatasubramanyan has spent over two decades applying a rare combination of quantitative skills, business knowledge, and the ability to think from first principles to real world business problems. He currently consults in several areas including supply chain and health care management. As a Fellow at i2 Technologies, he tackled supply chains problems in areas as diverse as computer assembly, semiconductor manufacturer, consumer goods, steel, and automotive. Prior to that, he worked with several airlines on their aircraft and crew scheduling problems. He topped off his days at IIT-Bombay and IIM-Ahmedabad with a Ph.D. in Operations Research from the University of Wisconsin-Madison.
He is presently based in Dallas, USA and travels extensively all over the world during the course of his consulting assignments. You can also find Narayan on Linkedin at: http://www.linkedin.com/in/narayan3rdeye