Online learning method in a decision system

#This invention relates to online learning in a decision system.

#Many decision systems are applied in a variety of dynamic situations where a system must adapt or learn its internal structure or parameters for each situation in order to optimize or maintain its performance. For example, an image-based decision system that inspects semiconductor wafers for defects or measures system alignment and registration precision can expect the image characteristics to change as new layers are added to the wafer during various processing stages. Other dynamic applications include live cell analysis that tracks a group of cells over time or patient specific medical image analysis (such as MRI, CT, X-ray) that needs to account for the differences between imaging system setups or patients' characteristics and response to the imaging system. There are also many non-image based dynamic decision system applications, including data mining and decision support in security and financial applications, speech recognition, mobile robots, medical diagnosis and treatment support, etc.

#Ideally, a learning process would take place in a supervised learning fashion wherein domain experts provide the truth labels associated with input data (labeled data). [Sergios Thodoridis and Konstantinos Koutroumbas, Pattern Recognition, Academic Press, 1999 pp 3-7.] This results in learning data for which the desired outcome has been provided and the system learns to match particular distributions of inputs to the associated outputs (desired response). In practice, however, such truth labeling presents significant problems in time, expense, and availability of data that often preclude its use. Furthermore, explicit learning after a system is placed into production interrupts normal system operation, which could significantly impact the productivity of the system.

#The preferred embodiments and other aspects of the invention will become apparent from the following detailed description of the invention when read in conjunction with the accompanying drawings which are provided for the purpose of describing embodiments of the invention and not for limiting same, in which:

#FIG. 1 shows a block diagram of a decision system processing flow, including startup and online learning processes;

#FIG. 2 shows a example diagram portraying densities of healthy and unhealthy patients as a function of systolic blood pressure;

#A kernel-based model has the flexibility to accurately represent a wide variety of distributions but requires a large amount of data in order to do so without a damaging amount of variance. If the sample set is too small, one model can vary widely from another model with both having the same number of samples (but different samples) from the same distribution. One embodiment of the invention combines the advantages of kernel-based and empirical modeling by choosing as atoms for the initial learning set a functionally derived set of atoms that constrains the kernel-based model and allows a limited number of shapes of the model distribution thereby assuring model stability even with a very small number of learning samples.

#An example is shown in FIG. 4 for a one-dimensional feature that has a model of the feature score distribution that is a step function. The feature score model is 4 f ⁡ ( x ) = ∑ i = 0 I 0 - 1 ⁢ ⁢ wg ⁡ ( x - i I ) + ∑ i = I 0 I - 1 ⁢ ⁢ rwg ⁡ ( x - i I ) where w = 1 I 0 + r ⁡ ( I - I 0 )

#for a model with I atoms, a transition from high-density to low density at 5 x 0 = I 0 I ,