With the rapid adoption of AI across domains, there are only a handful of fields that the technology is not a part of. Interestingly, the role of AI and data science in a particular field – operations research – is highly ambiguous. A common perception is that operation research (OR) is not useful for data science; furthermore, the overlap between data science and OR is misunderstood.
This misconception mainly stems from the marketing of OR products and services that are applied to the real world – quite often than not, the end-users do not have an understanding of the terms OR and data science. Another reason is that readily available machine learning models are available as packages of several platforms like Python and do not really contain specific OR models.
Sign up for your weekly dose of what's up in emerging technology.
In reality, however, OR problems are applicable to AI and data science. In fact, a lot of ideas in AI and data science problem solving have cross-pollinated from OR due to the large overlap in the techniques and methods used.
Speaking about the same was Rajeev Rajan, AVP, data science, Genpact, at the MLDS 2022 event.
Operation research and machine learning
Operation research is a fundamental part of the overall machine learning lifecycle. It is particularly useful when it comes to business problems that require parameter optimisation. Some of the examples of OR are:
- Enabling smart workforce management by forecasting resource requirements and optimising daily schedule for resources
- Increasing TV program viewership by optimal scheduling of programs’ promotion
- Enabling supply chain transformation by providing AI/ machine learning-based recommendations for optimised product utilisation
- AI-enabled forecasting for retail and eCommerce applications to optimise funnel and customer traffic
- Enabling data-driven optimisation for automated warehouse management, inspection and quality control
“We should think artificial intelligence, data science, and operations research as three different pieces. That said, even in machine learning, there is some type of optimisation, right, some type of mathematical modelling is going on, and we need to estimate the parameter right. As an ML guide or data scientist, one should not think that a basic operations research problem is outside their purview. If you may skip this step, you may not achieve the entire satisfaction that you were supposed to get out of bringing that particular ML or particular AI solutions for the customers,’’ explains Rajan.
Rajan explained operations research and its role in the overall machine learning scheme via a lot of examples. One of such use cases was about increasing the viewership of TV programs. The challenge here, as Rajan explained, is to apply machine learning techniques to prevent loss of revenue. This loss of revenue is observed as a result of a decline in viewership due to sub optimal scheduling of promotions between the programs.
To this end, Rajan proposes a solution where one collects data and executes mixed-integer programming in Python using Gurobi solver to generate an optimal promotion schedule. This results in a 4.5 per cent increase in revenue due to enhanced viewership of promoted shows.
An AI development lifecycle consists of the following steps:
- AIDLC kickoff: This step involves defining the problem to be solved.
- Ideate and assess: This step is about understanding the current state and accordingly defining the work scope.
- ML model development: Here, the machine learning solution is developed and tested.
- ML outputs given as OR inputs: Here, the OR techniques are used to make recommendations based on the outputs from the ML model. This is a critical step for the entire life cycle.
- Final step: Finally, the solution output is delivered to the client.
As per Rajan, following are some of the observations of incorporating OR in an AI initiative:
- Data science and OR are not always perceived as closely related. Most companies running AI and advanced analytics employ multidisciplinary teams that cover both.
- Hybrids of OR and data science techniques are effectively used in deploying end-to-end solutions.
- When the customers’ end goal is automating decision making, products are referred to as AI rather than OR.
- Using OR tools and techniques within AI applications will help in spreading large scale AI integrations.
- Developing OR as a skillset is a critical part of an effective AI initiative.