- What is ABAGAIL?
- How to get started with Abagail?
- How is ABAGAIL project organized?
- What are the different ways can I work with ABAGAIL?
- How to use ABAGAIL with Eclipse?
- Is there a tutorial or documentation for ABAGAIL?
- Can you generate graphs out of ABAGAIL library?
- How can I find out the actual number of iterations performed in Abagail?
- Could someone help explain the training_iterations in ABAGAIL? When I increase it, I get better accuracy.
- My neural network classifier is performing very poorly, how to fix it?
- How does the neural network get its weights updated?
- I am trying to understand what the count ones optimization function is doing, but I am not sure...
- TravelingSalesmanRouteEvaluationFunction - why the 1/total distance?
- My dataset has categorical features, can i use ABAGAIL?
- What are some characteristics of problems simulated annealing works well on?
- Getting build error: /home/unazi/ABAGAIL/build.xml:40: Unexpected attribute "additionalparam", how do I fix this?
ABAGAIL is an acronym of Absolute Best Andrew Guillory Artificial Intelligence Library.
from src > abagail.html:
This library is the result of close to a year of research and class work in artificial intelligence (AI). It contains a number of interconnected Java packages that implement machine learning and artificial intelligence algorithms. These are artificial intelligence algorithms implemented for the kind of people that like to implement algorithms themselves.
Start by download the source code from github You can use git to clone the repository:
git clone https://github.com/pushkar/ABAGAIL.git
or download the .zip file from github
After downloading the source code, decide how you want to work with the library, the most common options are listed here.
VERY BASIC INSTRUCTIONS FOR COMMAND LINE
- Install Java 8 SDK from here http://www.oracle.com/technetwork/java/javase/downloads/index.html
- Install Ant http://ant.apache.org/
- Clone or download source files from Git
- Go with command line to where the build.xml file is and run: ant (note: the ant executable should be in your path somehow if you installed ant correctly.. so will java and javac)
- Now run your scripts
At a high-level, the project structure is organized as follows:
dist
- Contains classes to generate distributions
func
- Contains superviced learning algorithms
- Contains clustering algorithms
opt
- Contains random search algorithms
- Contains optimization test functions
rl
- Contains reinforcement learning algorithms
shared
- Contains filters
- Contains IO and test running functions
util
- Contains utility methods and data structures
There are several options:
- use terminal or command line interface to compile and execute Java code
- use Java-compatible IDE with ABAGAIL as a library examples: Eclipse
- use Scala console with ABAGAIL as a library
- use Jython
-
fork
https://github.com/pushkar/ABAGAIL.git -
From the terminal clone your forked repo to a directory of your choice.
-
Create an Eclipse project, let's say it's "AbagailProject"
-
With the following steps you will convert the directory where you copied Abagail into an Eclipse Project
a. Go to File->New->Project.
b. On the menu double click on Java and select "Java Project from Existing Ant Build file". Click Next.
c. Browse to the Abagail folder and select the xml file. Click Finish.
- Now you need to include this new Eclipse project to the build folder of your project.
a. Right click on "AbagailProject" and select Properties.
b. On the left pane of the window select "Java Build Path."
c. Select the "Projects" tab.
d. Click Add...
e. Select the abagail project you just created (note that when you create the Eclipse project, the name of the project is "project" by default)
- To use any of the Abagail classes just use import. For instance, if you want to use the class ArffDataSetReader, you need to add the following import statement:
import shared.reader.ArffDataSetReader;
At this time the documentation is very limited. ABAGAIL comes with some documentation for you to look through when you build the project as well as a few examples. Users are actively encouraged to contribute to project documentation.
Currently no. There are several external libraries and software you can use to visualize results:
- Microsoft Excel
- R
- matplotlib
- JFreeChart
- matlab
- mathematica
- ...etc.
When you work with random search optimization problems, the number of iterations in the example problems is fixed.
For example:
FixedIterationTrainer fit = new FixedIterationTrainer(algorithm, 200000);`
will train the algorithm 200,000 times.
Alternatively, you can use ConvergenceTrainer which will train until convergence, or up to set number of maximum iterations, whichever occurs earlier:
ConvergenceTrainer trainer = new ConvergenceTrainer(algorithm);`
Could someone help explain the training_iterations in ABAGAIL? When I increase it, I get better accuracy.
In short, the more you train, the more accurate your weights become. This leads to better accuracy
Make sure your dataset values are normalized. See this article for more details.
Q:I was looking through the AbaloneTest.java code, and notice that each of the randomized algorithm has a train function that, in the case of SA for e.g., picks a neighbor and its fitness value (or sticks with the current point). However, I wasn't able to figure out how the NN gets updated with the new weights? I don't see anything where the train the SA code passes anything back to the concrete class of HillClimbingProblem.
Under ABAGAIL/src/func/nn/backprop/: BackPropagationNetworkFactory.java
private static BackPropagationNetworkFactory factory = new BackPropagationNetworkFactory();
/*********/
// returns instance of BackPropagationNetwork which has updateWeights method.
networks[i] = factory.createClassificationNetwork(new int[] {inputLayer, hiddenLayer, outputLayer});
/*********/
networks[i].setWeights(optimalInstance.getData());
Networks is an instance of BackPropagationNetwork.java, which has a method updateWeights(), which in turn extends class FeedForwardNetwork.java > LayeredNetwork.java > NeuralNetwork.java. The last one has setWeights() method.
Q: I am trying to understand what the count ones optimization function is doing. From the test class in ABAGAIL, I see that an array is created and filled up with all int=2. I think i understand the concept of counting all of the 1s in the vector, but I do not see how the array of all 2s turns into an array of 1s and 0s?
If you dig through the classes that uses the ranges variable, it will become clear. The 2 specifies how many different values are possible at any point in the vector (0 and 1).
The object of the Traveling Salesman problem is to minimize the distance in a route. However, the optimization algorithms maximize their fitness functions. So it returns the inverse of the distance, which gets larger as the distance gets smaller.
Returning the normal distance would result in the algorithms finding the longest routes, which isn't what we're interested in.
Convert the categorical features to binary, e.g., if you have Country as a variable, then each value of country will have its own column with 1 for the country that's present, and 0 for values that isn't.
Getting ant build error: /home/unazi/ABAGAIL/build.xml:40: Unexpected attribute "additionalparam", how do I fix this?
Make sure you are using a version where this issue has been fixed. See: #31
Generally, simulated annealing works well if there are multiple minimum and you need to find the global minimum because SA will (hopefully) break out of the local minimum.
in the ABAGAIL library look for sample test problems under opt > example directory