Thank you for answering. I went through both the files. First one is more of probabilistic technique based work. That too it does not give a detailed description of the procedure. The second one is related to business analytics.
1) I am looking for some standards works IEEE Transactions/ Patents.
2) With more inclination towards data acquisition/ data processing/ machine learning/ pattern recognition/ event forecasting.
3) Basically, I am referring to signal & image processing based applications.
Thank you for your suggestion. Hopefully this book will address the modeling part of my query. Any more insights on suitable forecasting techniques in such a scenario?
I believe that any particular suggestion depends on your scenario, e.g. do you have the ground truth output of your stream? In the case of an affirmative answer, you could follow a completely different branch of solutions.
What's your main goal (time, precision or memory saving)?
I think you need to define this type of constraints, afterwards setting up a pipeline of techniques.
The priority order would be precision> time> memory. Actually, I need to predict the behavior of a continuously changing pattern based on machine learning. Suppose I have consecutive images or numerical data corresponding to a pattern, with respect to time. Using this data I need to predict the pattern orientation/ behavior in the next few seconds. Do you think reinforcement learning would be suitable for this problem? Any other suggestions are also welcome.
1) Considering that you are handling an image stream (like video), you need to pay attention to the precision and time trade-off. I mean, the features you are extracting could compromise your time. On the other hand a low complexity method could not reach your desired precision.
2) If the whole combination of orientation/behavior are previously known you could use a simple decision technique, not necessarily a machine learning algorithm to build a model. Just apply a technique as ADWIN [1] to tackle the stream constraints.
3) Reinforcement and Active Learning are very good solutions, but you need to think about the availability of specialist and model changing (scenarios with concept-drift). Otherwise, Some approaches as VFDT [1] requests a "Grace Period" to online build a model capable of deal with your problem (without new concepts or drifts).