Estimating Link Signatures with Machine Learning Algorithms
Machine Learning Algorithms: A Brief Introduction Supervised Learning, Unsupervised Learning, Reinforcement Classification vs Regression Problems
Supervised Learning, Unsupervised Learning, and Reinforcement Supervised: specifically labeled training data Unsupervised: data not marked, machine learns on the fly Reinforcement: machine attempts to get the largest reward possible
Classification vs Regression Supervised learning includes two categories of algorithms: a)Classification: for categorical response values, where the data can be separated into specific “classes” examples: ham/samp , image of a cat/ not a image of a cat, etc. b)Regression: for continuous-response values(number) example: Housing Prices or tomorrow’s temperature,etc.
Defining our Problem in the Context of Machine Learning A Formal Definition of Machine Learning: A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E. Tom M. Mitchell Task (T): Estimate link signatures between transmitters and receivers Experience (E): measurements of link signatures in the given area for other nodes that are not the given transmitter receiver channel Performance (P): Correctness or successfully guessing the link signature
Different Algorithms to Try Support Vector Machine (SVM) --- very promising! Regression based (SVR) Decision Tree / Regression Tree Random Forest (a cluster of decision trees)
Support Vector Machines (SVM) Advantages Effective in high dimensional spaces. Prediction accuracy is generally high Memory Efficient(support vector, training point) Disadvantages Long training time Difficult to understand the learned function(weights)
Decision Tree Algorithms Advantages Simple, relatively easy to understand Data does not need to be linearly separable Not “black box” Disadvantages Tend to overfit to training data Instability -- if the data is even a little different than expected, you get a different tree
Features used in Our Problem Our assumptions about what can be used to predict the next value Previous features used in neural network approach New features we may consider adding?
Work for the Upcoming Week Implement SVR and Decision Tree Algorithms in Matlab Try to come up with new features to give to the algorithms Look at more algorithms to potentially use