Learning and optimizing your business: the LION way.
Learning and Intelligent OptimizatioN (LION).
The two tasks of learning models and determining optimal configurations often must proceed hand in hand to solve
challenging business problems.
The following example is schematic, to make things more concrete you may want to concentrate on your current business.
For sure, you want to maximize some objectives (like profit, or quality, ...), but using software
to help in the process is challenging if the software requires that you specify your
business model before starting . This is the main reason why computers are not used a lot -- they are
not used as they could be used --
to improve business processes and decisions.
LION (Learning and Intelligent OptimizatioN) advocates a radically different path. LION does not ask you
to specify a model, but to experiment with the current system. The appropriate model will be learnt by
the software and then used to identify a better solution, in a learning cycle with more than one
iteration. Good news: you do not need a super-expert in mathematics or in computer science to improve
LION in action!
If you do not like words, you can see an example of our LIONoso in action in the following YouTube movie (7 minutes).
If you prefer words, here you are. Our best suggestion is to experiment with your business problem (by downloading the
evaluation version of LIONoso). In the meantime, you can get an idea about how it works from the following abstract.
The objective is to find the minimum value of an external system. What you have to do is (and you may want to keep the above figure as a visual reference):
1. Connect your business to LIONoso. In this demo, your business is an external software module called "Bells" described
in the Scilab framework. In the movie, the external system was an electro-mechanical scanning equipment. What you attach to LION is your business.
How to setup a quick "learning and optimization" cycle is our business.
2. Experiment with your system by creating inputs and measuring outputs (DOE module in the picture). NOTE: your system can be a complete "black box"
at the beginning. The results of your experiments will look like the following picture (where the output value obtained is plotted as
a function of the input values).
3. Have the software build a flexible model explaining how the output depends on the inputs (a polynomial model of degree 5 is created in this case).
4. Instruct LIONoso to determine an optimal solution by using the model. Take the optimal inputs and pass them though the original "Bells" system.
The newly determined configurations reach lower values. Good, we made some progress (see the pale balls generated in the plot below, with a value f(x,y) close to -2.0)
5. Setup an automated learning-optimizing cycle by connecting the optimizer (RSO) to the external system ("Bells"). Now when a better point is determined by
using the surrogate model, it will be added to the set of experiments, which in turn will be used to create a better model, which in turn will be optimized, ...
You decide when to stop the iterations, for example when the incremental improvement is negligible in the latest iteration, or when your time for
experimenting is finished and it is time to act.
By the way, if you are curious, this exact external demo system is characterized by a double-bell shape when plotting the output values as a
function of the input values. Indeed, LIONoso managed to identify the optimal solution, and it did so without asking you
for a mathematical definition of your system.
References: The LION way
Roberto Battiti and Mauro Brunato. LIONlab, University of Trento, Feb 2014.