I suggest you to use ANN or ANFIS. lt is clear that you can also use any algorithm to see which one is better.

I suggest you to use ANN or ANFIS. lt is clear that you can also use any algorithm to see which one is better.

As I understand your question, you are looking for a technique to generate one or more models according to the huge data set, to enforce the responsetime for a lookup in that model.

That huge dataset also can be considered as a model. To reduce that model you have to eliminate irrelevant information and group environment descriptions (demand) which result in the same action (cycle length and lane group)

A sophisticated approach might be to train a Learning-Classifier-System (LCS), e.g. the Zeroth-level Classifier System (ZCS) from Wilson (1994). After training, the Classifier-Set will be the model you need. The LCS is also used by the Organic Computing Group to control traffic lights in a self-organizing manner.

You also might use a self-organizing feature map (SOM). Often it is used for reducing the data dimension und for clustering approaches.

(A)

Yoo, C., Ramirez, L., & Liuzzi, J. (2014). Big data analysis using modern statistical and machine learning methods in medicine. International neurourology journal, 18(2), 50-57.‏

introduce modern statistical machine learning and bioinformatics approaches that have been used in learning statistical relationships from big data in medicine and behavioral science. Especially they report on Bayesian networks as an appropriate ML method for Big Data.

(B)

Suthaharan, S. (2014). Big Data Classification: Problems and Challenges in Network Intrusion Prediction with Machine Learning.‏

reports on three major problems that make the ML techniques unsuitable for solving Big data classification problems are: (1) An ML technique that is trained on a particular labeled datasets or data domain may not be suitable for another dataset or data domain – that the classification may not be robust over different datasets or data domains; (2) An ML technique is in general trained using a certain number of class types and hence a large varieties of class types found in a dynamically growing dataset will lead to inaccurate classification results; and (3) An ML technique is developed based on a single learning task, and thus they are not suitable for today’s multiple learning tasks and knowledge transfer requirements of Big Data analytics.

Suthaharan  wrote that SVM's variants, for example, are not suitable for Big Data analytics due to their high computational complexity.

Suthaharan suggests two ML methods that are possible suitable for big data: representation-learning (CDRL) and unit-circle algorithm (UCA) [15] .

(C)

Additional relevant papers are:

Ratner, B. (2011). Statistical and machine-learning data mining: techniques for better predictive modeling and analysis of big data. CRC Press.‏

Condie, T., Mineiro, P., Polyzotis, N., & Weimer, M. (2013, June). Machine learning for big data. In Proceedings of the 2013 ACM SIGMOD International Conference on Management of Data (pp. 939-942). ACM.‏

Rudin, C., & Vahn, G. Y. (2014). The big data newsvendor: Practical insights from machine learning.‏

Ma, C., Zhang, H. H., & Wang, X. (2014). Machine learning for Big Data analytics in plants. Trends in plant science, 19(12), 798-808.‏

Rudin, C., & Vahn, G. Y. (2014). The big data newsvendor: Practical insights from machine learning.‏

Unfortunately, we have found serious tradeoffs in terms of quality and speed when selecting optimal underlying ML mechanisms.  For long term time-series prediction, for example, nothing beats SVM-based methods -- as pointed out above they are SLOW, but not exponentially so, so they can run assuming you have the computational power and *some time* e.g. (don't need it real-time to the millisecond etc).  What is key is pre-processing the inputs to the SVMs so they aren't dealing with too much raw data - CDRL is really a form of that as I understand it.  Again, I'm just talking about time-series prediction using vast medical data on large patient populations where the object is to predict what will happen to a given patient X from t(x) to t(x+n) where you have patient data from t1 to t(present) and you also have thousands of other similar patient records.  So, this may not be terribly relevant to your traffic application.  Also, we need to be "near real time" meaning we can wait at least 10 seconds or so for a prediction; we don't need an "instantaneous" calculation.  Perhaps CDRL is best angle for you, but beware speed-quality tradeoffs.   

There is no AI technique which is superior to all the others over the universe of problems - this is the well-known "No Free Lunch" theorem. See http://www.no-free-lunch.org/    You will need to do some preliminary data analysis on your dataset then some experiments to find promising candidates.

As others have said, there is no single best approach. Also, you need to quanitify 'huge" as one person's "huge" might be another person's "moderately-sized."  A starting point might be this graphic. It was developed for a particular toolkit (scikit learn), but is generally applicable. The one major family of techniques missing is neural nets, becasue scikit learn doesn't have a general neural net set of tools.

http://scikit-learn.org/stable/tutorial/machine_learning_map/