Bungee Mining: take the plunge . . .

There are two phases to the Bungee Mining approach: the first phase - the jump - is characterized by a rapid exploration of the data, intellectual courage, and a sense of freedom, exploration and creativity. The last phase - the elastic oscillations - is characterized by rapid adaptation of the first results to more detailed constraints and final user requirements, with possible quick trajectory changes, but with methods ensuring rapid convergence to the final objective and to the exploitation of the obtained insight.

Methods to achieve these objectives are based on interaction, mutual learning and feedback loops both from the decision maker to the software system and from the software system to the decision maker. The end user discovers possibilities and may change his objectives after the first results are available. The system saves the user's judgments and responses and uses them in an optimal way to focus his search more and more towards his real preferences. A quick reaction based on a sound machine learning approach is coupled to proactive actions to prepare for changes, measure progress and come progressively closer to the target.

No actual programming is required for many basic application scenarios, when performance measures and objectives are already available, apart from connecting the data mining and visualization routines to the data and the performance measures. This activity is usually rapid and seamless, provided that modern tools, programming languages and databases have been used in the system modeling and data collection tasks. The Bungee Mining team at Grapheur.com is available for suggestions about how the approach can be deployed for your application case.

Standard operations research and problem solving method assume that the criterion to be optimized is given, and actually is a computable function of some clearly defined input parameters. Unfortunately the real-world is not a math classroom and defining the objectives of the problem solving / optimization effort is the pain point which often consumes most resources, time, and money. Try asking a manager "Please tell me the function that you are optimizing", responses may range from laughter to anger. Recognizing this issue changes the computer science effort from improving by 0.1% the results on artificially generated classroom problems to devoting most computational resources to help a user learning more about his problem, expressing his explicit and implicit preferences, identifying methods, models and parameters in a semi-automated manner. You may call it "computer-supported consultancy" if you want.

In business, technical and scientific processes, the raw data are only the starting point for any investigation, design or problem solving effort. Real insight can be obtained only through models intended to "explain" the raw data in a concise and summarized manner. This requires a spiraling process where models are identified, learnt, validated or improved though various feedback paths involving the designer or careful tests with additional data if appropriates. This "on the job learning" method is at the foundation of Reactive Search Optimization, advocating learning loops to improve the results while a system is working on a specific case. Automation is increased by the use of adaptation and machine learning schemes during the iterations, relieving the end user from the expensive and slow tuning process required for solving the case of interest.

In multi-objective optimization many measures (success criteria) are available to the designer or decision maker and a comprehensive combination of the measures into a single criterion is not possible, because of either lack of knowledge, excessive costs, or a desire to adapt the preferences to preliminary results. Let's consider planning for a vacation. You look for a cheap and good quality hotel and you are prepared to deal with compromises: cheaper hotels are typically related to poorer quality, at least in a free market. Imagine that quality is measured by the number of stars, as it is in many European countries. You are reluctant to answer quantitative questions at the beginning, such as "How much are you willing to pay for one more quality star?", or "By how much is quality more important than price for you?". This does not mean that you are not capable of finding preferred solutions, but that these solutions have to be found by iterations, where some possibilities are shown, judged, and the feedback acquired by the support system (maybe a tourist agency in this case) to search for better and better solutions according to your explicit and implicit preferences. In more complex practical applications, the number of possible solutions can be very large, and visualization comes to the rescue to help develop insight about the problem, identify prototypes via clustering techniques, and devise proper ways of looking at the original data.

When embodied into the Grapheur product, Reactive Search Optimization (RSO) coupled with Bungee Mining works as follows: Free your mind from software complications (choices, tuning, etc.) and concentrate your creativity on exploring relationships in your data, on specifying your preferences, and on designing your preferred solution. Do not let the tool bother you, let your preferences and insight come into the spotlight. Navigate in the solution space to identify interesting patterns and promising alternatives, focus onto the most interesting parts, develop a mental map of the possibilities. Build models to predict the output for new cases, visualize how the various input data influence the output. Develop trailblazing dynamic visualizations that go beyond the standard and static plots. Next steps:
—> consider downloading the free evaluation software from the Grapheur website
—> buy a license if satisfied (basic packages start from less than 1000 Euros)
—> contact Reactive Search for consultancy if you have a challenging problem to address.