JeanFrancoisPuget 2700028FGP Visits (2615)
Decision management is a hot topic these days. At IBM it is defined as a combination of predictive analytics and business rules. Used separately or in combination, these analytics techniques provide a powerful way to automate, optimize, and govern repeatable business decisions. Readers interested in learning more about decision management should have a look at Jame's Taylor's site in addition to IBM's decision management pages.
Wait a minute. We just wrote three sentences about decision making, didn't we? Isn't mathematical optimization all about making better decisions?
How can we relate optimization and decision management? Rather than going on with an abstract discussion, let me use an example that shows how optimization can complement a decision management solution.
That example is based on something we successfully delivered few years ago to a large retail bank. The problem for that bank was to select the next commercial actions to be performed. Should someone call customer A to discuss a loan offer? Should someone else send an email to customer B about the new Visa card of the bank? Should someone else send a leaflet about housing insurance to customer C? When you have millions of customers the selection of these actions can be complex. One has to deal with various limits (budgets, available time from available agents to perform the actions, etc). More importantly, one has to rate actions. Which actions are most likely to yield revenue? Which actions are eligible for a given customer?
Decision management techniques are the tool of choice for answering the last two questions. Here is how it was done for that bank.
For each customer perform these steps:
That process is repeated for each customer.
This is a quite powerful and generic process. It is so generic that it has been packaged in IBM Next Best Action offering. I recommend this Next Best Action technical description to interested readers. It expands and details the various steps I sketch above.
While it is a quite useful process, it has one major drawback. Issue is that it provides a myopic view that may not lead to a good use of resources overall. Let me show this with a small example.
We have two possible actions, plan 1, and plan 2. The outcome for plan 1 is 65 in case of success, and the cost is 20. The outcome form plan 2 is 50 and the cost is 15.
Then we have a total budget of 35. I am using budget but we could use several limits such as number of available agents for calling, number of available agents for on site visits, etc. I only use one limit, called budget here, for the sake of simplicity.
Assume further that there are 3 customers, named A, B, and C. The table below shows for each customer (col 1) the eligible actions (col 2), the outcome in case of success (col 3), the likelihood of success for that action and that customer (col 4), the expected outcome (col 5).
Then col 6 shows what next best action would do:
Let us look at a global optimization process. For it we compute all possible expected outcome. We have already computed it for customer B with plan 1 and plan 2, and we have computed it for customer C with plan 2. The remaining one, ie the expected outcome for customer C and plan 2 is 65 x 50% = 33
Then the problem is to select the best set of actions, ie the set of actions whose cost is within budget limit, and whose expected outcome is the largest. The answer is to propose plan 2 to customer B and plan 1 to customer C. This way the expected outcome is 25 + 33 = 58, which is way better than what we got when dealing with customers one at a time.
This example is not artificial. As a matter of fact is could be hard to construct real examples where looking at each decision in turn would lead to a very good use of resources overall.
We have developed, and deployed, an optimization solution for the selection of the best commercial action for that large bank. That solution was able to select 100k actions from a pool of 1M actions in a few hours. When volumes are much smaller, say hundreds of persons, and thousands of possible response, then compute time goes down to minutes, if not seconds. The use of large scale optimization enabled the bank to meet its operational constraint while maximizing revenue generation.
This example show that optimization can complement the use of decision management techniques. Optimization improves the overall efficiency by looking at the whole picture, while decision management techniques are used to process the actions for a given customer.