I am trying to write an adaptive controller for a control system, namely a power management system using Q-learning. I recently implemented a toy RL problem for the cart-pole system and worked out the formulation of the helicopter control problem from Andrew NG's notes. I appreciate how value function approximation is imperative in such situations. However both these popular examples have very small number of possible discrete actions. I have three questions:
1) What is the correct way to handle such problems if you don't have a small number of discrete actions? The dimensionality of my actions and states seems to have blown up and the learning looks very poor, which brings me to my next question.
2) How do I measure the performance of my agent? Since the reward changes in conjunction with the dynamic environment, at every time-step I can't decide the performance metrics for my continuous RL agent. Also unlike gridworld problems, I can't check the Q-value table due to huge state-action pairs, how do I know my actions are optimal?
3) Since I have a model for the evoluation of states through time. States = [Y, U]. Y[t+1] = aY[t] + bA, where A is an action.
Choosing discretization step for actions A will also affect how finely I have to discretize my state variable Y. How do I choose my discretization steps?
Thanks a lot!
You may use a continuous action reinforcement learning algorithm and completely avoid the discretization issue. I'd suggest you to take a look at CACLA.
As for the performance, you need to measure your agent's accumulated reward during an episode with learning turned off. Since your environment is stochastic, take many measurements and average them.
Have a look at policy search algorithms. Basically, they directly learn a parametric policy without an explicit value function, thus avoiding the problem of approximating the Q-function for continuous actions (eg, no discretization of the action space is needed).
One of the easiest and earliest policy search algorithm is policy gradient. Have a look here for a quick survey about the topic. And here for a survey about policy search (currently, there are more recent techniques, but that's a very good starting point).
In the case of control problem, there is a very simple toy task you can look, the Linear Quadratic Gaussian Regulator (LQG). Here you can find a lecture including this example and also an introduction to policy search and policy gradient.
Regarding your second point, if your environment is dynamic (that is, the reward function of the transition function (or both) change through time), then you need to look at non-stationary policies. That's typically a much more challenging problem in RL.
Background: I am trying to find a list of floating point parameters for a low level controller that will lead to balance of a robot while it is walking.
Question: Can anybody recommend me any local search algorithms that will perform well for the domain I just described? The main criteria for me is the speed of convergence to the right solution.
Any help will be greatly appreciated!
P.S. Also, I conducted some research and found out that "Evolutianry
Strategy" algorithms are a good fit for continuous state space. However, I am not entirely sure, if they will fit well my particular problem.
More info: I am trying to optimize 8 parameters (although it is possible for me to reduce the number of parameters to 4). I do have a simulator and a criteria for me is speed in number of trials because simulation resets are costly (take 10-15 seconds on average).
One of the best local search algorithms for low number of dimensions (up to about 10 or so) is the Nelder-Mead simplex method. By the way, it is used as the default optimizer in MATLAB's fminsearch function. I personally used this method for finding parameters of some textbook 2nd or 3rd degree dynamic system (though very simple one).
Other option are the already mentioned evolutionary strategies. Currently the best one is the Covariance Matrix Adaption ES, or CMA-ES. There are variations to this algorithm, e.g. BI-POP CMA-ES etc. that are probably better than the vanilla version.
You just have to try what works best for you.
In addition to evolutionary algorithm, I recommend you also check reinforcement learning.
The right method depends a lot on the details of your problem. How many parameters? Do you have a simulator? Do you work in simulation only, or also with real hardware? Speed is in number of trials, or CPU time?
Consider an optimization problem of some dimension n, Given some linear set of equations(inequalities) or constraints on the inputs which form a convex region, finding the maximum\minimum value of some expression which is some linear combination of inputs(or dimensions).
For larger dimension, these optimization problems take much time to give the exact answer.
So, can we use machine learning techniques, to get some approximate solution in lesser time.
if we can use machine learning techniques in this context, How the Training set should be??
Do you mean "How big should the training set be?" If so, then that is very much a "how long is a piece of string" question. It needs to be large enough for the algorithm being used, and to represent the data that is being modeled.
This doesn't strike me as being especially focused on machine learning, as is typically meant by the term anyway. It's just a straightforward constrained optimization problem. You say that it takes too long to find solutions now, but you don't mention how you're trying to solve the problem.
The simplex algorithm is designed for this sort of problem, but it's exponential in the worst case. Is that what you're trying that's taking too long? If so, there are tons of metaheuristics that might perform well. Tabu search, simulated annealing, evolutionary algorithms, variable depth search, even simple multistart hill climbers. I would probably try something along those lines before I tried anything exotic.
I am working on testing several Machine Learning algorithm implementations, checking whether they can work as efficient as described in the papers and making sure they could offer a great power to our statistic NLP (Natural Language Processing) platform.
Could u guys show me some methods for testing an algorithm implementation?
1)What aspects?
2)How?
3)Do I have to follow some basic steps?
4)Do I have to consider diversity specific situations when using different programming languages?
5)Do I have to understand the algorithm? I mean, does it offer any help if I really know what the algorithm is and how it works?
Basically, we r using C or C++ to implement the algorithm and our working env is Linux/Unix. Our testing methods only focus on black box testing and testing input/output of functions. I am eager to improve them but I dont have any better idea now...
Great Thx!! LOL
For many machine learning and statistical classification tasks, the standard metric for measuring quality is Precision and Recall. Most published algorithms will make some kind of claim about these metrics, or you could implement them and run these tests yourself. This should provide a good indicative measure of the quality you can expect.
When you talk about efficiency of an algorithm, this is usually some statement about the time or space performance of an algorithm in terms of the size or complexity of its input (often expressed in Big O notation). Most published algorithms will report an upper bound on the time and space characteristics of the algorithm. You can use that as a comparative indicator, although you need to know a little bit about computational complexity in order to make sure you're not fooling yourself. You could also possibly derive this information from manual inspection of program code, but it's probably not necessary, because this information is almost always published along with the algorithm.
Finally, understanding the algorithm is always a good idea. It makes it easier to know what you need to do as a user of that algorithm to ensure you're getting the best possible results (and indeed to know whether the results you are getting are sensible or not), and it will allow you to apply quality measures such as those I suggested in the first paragraph of this answer.
A developer I am working with is developing a program that analyzes images of pavement to find cracks in the pavement. For every crack his program finds, it produces an entry in a file that tells me which pixels make up that particular crack. There are two problems with his software though:
1) It produces several false positives
2) If he finds a crack, he only finds small sections of it and denotes those sections as being separate cracks.
My job is to write software that will read this data, analyze it, and tell the difference between false-positives and actual cracks. I also need to determine how to group together all the small sections of a crack as one.
I have tried various ways of filtering the data to eliminate false-positives, and have been using neural networks to a limited degree of success to group cracks together. I understand there will be error, but as of now, there is just too much error. Does anyone have any insight for a non-AI expert as to the best way to accomplish my task or learn more about it? What kinds of books should I read, or what kind of classes should I take?
EDIT My question is more about how to notice patterns in my coworker's data and identify those patterns as actual cracks. It's the higher-level logic that I'm concerned with, not so much the low-level logic.
EDIT In all actuality, it would take AT LEAST 20 sample images to give an accurate representation of the data I'm working with. It varies a lot. But I do have a sample here, here, and here. These images have already been processed by my coworker's process. The red, blue, and green data is what I have to classify (red stands for dark crack, blue stands for light crack, and green stands for a wide/sealed crack).
In addition to the useful comments about image processing, it also sounds like you're dealing with a clustering problem.
Clustering algorithms come from the machine learning literature, specifically unsupervised learning. As the name implies, the basic idea is to try to identify natural clusters of data points within some large set of data.
For example, the picture below shows how a clustering algorithm might group a bunch of points into 7 clusters (indicated by circles and color):
(source: natekohl.net)
In your case, a clustering algorithm would attempt to repeatedly merge small cracks to form larger cracks, until some stopping criteria is met. The end result would be a smaller set of joined cracks. Of course, cracks are a little different than two-dimensional points -- part of the trick in getting a clustering algorithm to work here will be defining a useful distance metric between two cracks.
Popular clustering algorithms include k-means clustering (demo) and hierarchical clustering. That second link also has a nice step-by-step explanation of how k-means works.
EDIT: This paper by some engineers at Phillips looks relevant to what you're trying to do:
Chenn-Jung Huang, Chua-Chin Wang, Chi-Feng Wu, "Image Processing Techniques for Wafer Defect Cluster Identification," IEEE Design and Test of Computers, vol. 19, no. 2, pp. 44-48, March/April, 2002.
They're doing a visual inspection for defects on silicon wafers, and use a median filter to remove noise before using a nearest-neighbor clustering algorithm to detect the defects.
Here are some related papers/books that they cite that might be useful:
M. Taubenlatt and J. Batchelder, “Patterned Wafer Inspection Using Spatial Filtering for Cluster Environment,” Applied Optics, vol. 31, no. 17, June 1992, pp. 3354-3362.
F.-L. Chen and S.-F. Liu, “A Neural-Network Approach to Recognize Defect Spatial Pattern in Semiconductor Fabrication.” IEEE Trans. Semiconductor Manufacturing, vol. 13, no. 3, Aug. 2000, pp. 366-373.
G. Earl, R. Johnsonbaugh, and S. Jost, Pattern Recognition and Image Analysis, Prentice Hall, Upper Saddle River, N.J., 1996.
Your problem falls in the very broad field of image classification. These types of problems can be notoriously difficult, and at the end of the day, solving them is an art. You must exploit every piece of knowledge you have about the problem domain to make it tractable.
One fundamental issue is normalization. You want to have similarly classified objects to be as similar as possible in their data representation. For example, if you have an image of the cracks, do all images have the same orientation? If not, then rotating the image may help in your classification. Similarly, scaling and translation (refer to this)
You also want to remove as much irrelevant data as possible from your training sets. Rather than directly working on the image, perhaps you could use edge extraction (for example Canny edge detection). This will remove all the 'noise' from the image, leaving only the edges. The exercise is then reduced to identifying which edges are the cracks and which are the natural pavement.
If you want to fast track to a solution then I suggest you first try the your luck with a Convolutional Neural Net, which can perform pretty good image classification with a minimum of preprocessing and noramlization. Its pretty well known in handwriting recognition, and might be just right for what you're doing.
I'm a bit confused by the way you've chosen to break down the problem. If your coworker isn't identifying complete cracks, and that's the spec, then that makes it your problem. But if you manage to stitch all the cracks together, and avoid his false positives, then haven't you just done his job?
That aside, I think this is an edge detection problem rather than a classification problem. If the edge detector is good, then your issues go away.
If you are still set on classification, then you are going to need a training set with known answers, since you need a way to quantify what differentiates a false positive from a real crack. However I still think it is unlikely that your classifier will be able to connect the cracks, since these are specific to each individual paving slab.
I have to agree with ire_and_curses, once you dive into the realm of edge detection to patch your co-developers crack detection, and remove his false positives, it seems as if you would be doing his job. If you can patch what his software did not detect, and remove his false positives around what he has given you. It seems like you would be able to do this for the full image.
If the spec is for him to detect the cracks, and you classify them, then it's his job to do the edge detection and remove false positives. And your job to take what he has given you and classify what type of crack it is. If you have to do edge detection to do that, then it sounds like you are not far from putting your co-developer out of work.
There are some very good answers here. But if you are unable to solve the problem, you may consider Mechanical Turk. In some cases it can be very cost-effective for stubborn problems. I know people who use it for all kinds of things like this (verification that a human can do easily but proves hard to code).
https://www.mturk.com/mturk/welcome
I am no expert by any means, but try looking at Haar Cascades. You may also wish to experiment with the OpenCV toolkit. These two things together do face detection and other object-detection tasks.
You may have to do "training" to develop a Haar Cascade for cracks in pavement.
What’s the best approach to recognize patterns in data, and what’s the best way to learn more on the topic?
The best approach is to study pattern recognition and machine learning. I would start with Duda's Pattern Classification and use Bishop's Pattern Recognition and Machine Learning as reference. It would take a good while for the material to sink in, but getting basic sense of pattern recognition and major approaches of classification problem should give you the direction. I can sit here and make some assumptions about your data, but honestly you probably have the best idea about the data set since you've been dealing with it more than anyone. Some of the useful technique for instance could be support vector machine and boosting.
Edit: An interesting application of boosting is real-time face detection. See Viola/Jones's Rapid Object Detection using a Boosted Cascade of Simple
Features (pdf). Also, looking at the sample images, I'd say you should try improving the edge detection a bit. Maybe smoothing the image with Gaussian and running more aggressive edge detection can increase detection of smaller cracks.
I suggest you pick up any image processing textbook and read on the subject.
Particularly, you might be interested in Morphological Operations like Dilation and Erosion, which complements the job of an edge detector. Plenty of materials on the net...
This is an image processing problem. There are lots of books written on the subject, and much of the material in these books will go beyond a line-detection problem like this. Here is the outline of one technique that would work for the problem.
When you find a crack, you find some pixels that make up the crack. Edge detection filters or other edge detection methods can be used for this.
Start with one (any) pixel in a crack, then "follow" it to make a multipoint line out of the crack -- save the points that make up the line. You can remove some intermediate points if they lie close to a straight line. Do this with all the crack pixels. If you have a star-shaped crack, don't worry about it. Just follow the pixels in one (or two) directions to make up a line, then remove these pixels from the set of crack pixels. The other legs of the star will recognized as separate lines (for now).
You might perform some thinning on the crack pixels before step 1. In other words, check the neighbors of the pixels, and if there are too many then ignore that pixel. (This is a simplification -- you can find several algorithms for this.) Another preprocessing step might be to remove all the lines that are too thin or two faint. This might help with the false positives.
Now you have a lot of short, multipoint lines. For the endpoints of each line, find the nearest line. If the lines are within a tolerance, then "connect" the lines -- link them or add them to the same structure or array. This way, you can connect the close cracks, which would likely be the same crack in the concrete.
It seems like no matter the algorithm, some parameter adjustment will be necessary for good performance. Write it so it's easy to make minor changes in things like intensity thresholds, minimum and maximum thickness, etc.
Depending on the usage environment, you might want to allow user judgement do determine the questionable cases, and/or allow a user to review the all the cracks and click to combine, split or remove detected cracks.
You got some very good answer, esp. #Nate's, and all the links and books suggested are worthwhile. However, I'm surprised nobody suggested the one book that would have been my top pick -- O'Reilly's Programming Collective Intelligence. The title may not seem germane to your question, but, believe me, the contents are: one of the most practical, programmer-oriented coverage of data mining and "machine learning" I've ever seen. Give it a spin!-)
It sounds a little like a problem there is in Rock Mechanics, where there are joints in a rock mass and these joints have to be grouped into 'sets' by orientation, length and other properties. In this instance one method that works well is clustering, although classical K-means does seem to have a few problems which I have addressed in the past using a genetic algorithm to run the interative solution.
In this instance I suspect it might not work quite the same way. In this case I suspect that you need to create your groups to start with i.e. longitudinal, transverse etc. and define exactly what the behviour of each group is i.e. can a single longitudinal crack branch part way along it's length, and if it does what does that do to it's classification.
Once you have that then for each crack, I would generate a random crack or pattern of cracks based on the classification you have created. You can then use something like a least squares approach to see how closely the crack you are checking fits against the random crack / cracks you have generated. You can repeat this analysis many times in the manner of a Monte-Carlo analysis to identify which of the randomly generated crack / cracks best fits the one you are checking.
To then deal with the false positives you will need to create a pattern for each of the different types of false positives i.e. the edge of a kerb is a straight line. You will then be able to run the analysis picking out which is the most likely group for each crack you analyse.
Finally, you will need to 'tweak' the definition of different crack types to try and get a better result. I guess this could either use an automated approach or a manual approach depending on how you define your different crack types.
One other modification that sometimes helps when I'm doing problems like this is to have a random group. By tweaking the sensitivity of a random group i.e. how more or less likely a crack is to be included in the random group, you can sometimes adjust the sensitivty of the model to complex patterns that don't really fit anywhere.
Good luck, looks to me like you have a real challenge.
You should read about data mining, specially pattern mining.
Data mining is the process of extracting patterns from data. As more data are gathered, with the amount of data doubling every three years, data mining is becoming an increasingly important tool to transform these data into information. It is commonly used in a wide range of profiling practices, such as marketing, surveillance, fraud detection and scientific discovery.
A good book on the subject is Data Mining: Practical Machine Learning Tools and Techniques
(source: waikato.ac.nz) ](http://www.amazon.com/Data-Mining-Ian-H-Witten/dp/3446215336 "ISBN 0-12-088407-0")
Basically what you have to do is apply statistical tools and methodologies to your datasets. The most used comparison methodologies are Student's t-test and the Chi squared test, to see if two unrelated variables are related with some confidence.