Here you will find updates on my Master's Thesis project named Rantbot. The specification can be found here (PDF).

It has been too long since my last update here. I have since continued working on the report and have gotten a good way into the first draft. I have also started my new job as a Systems Developer for Lehman Brothers, so I will only be able to work on my thesis during the evenings (and weekends). The next thing on my list is to fix up the code a bit for the trainer, so that I can run some final tests and document them.

I apologize for the lack of updates recently, but I have mostly been working on the report so there has not been that much to post about. On Monday I had a meeting with my supervisor to discuss some problems I had run into while doing experiments, and I am now working to solve those problems.

It has been a while since my last update, and since then I have done some experiments and tweaked the trainer a bit. I have now started writing more in my report, and as it looks now I will be done sometime in February.

On Monday I had another meeting with my supervisor where I showed my GUI and we came up with a todo list for me, as I am done with the implementation part and will start working on experiments and documentation (after some minor adjustments to my program). Today I have created a command line program in Lisp which converts a file of categories and sentences into a network, and then saves that file in a format used by my training application (the GUI).

I have now put all the functionality into the GUI, there is a screenshot of it in action here. I have added the sentencecs I have come up with (125 of them) to a network and started training on that network. Tomorrow I will continue training and I might also have a meeting with my supervisor, now that I finally have something to show again.

There is only a little bit left now in the functionality of the GUI, so I will be done with it tomorrow. I am looking forward to trying it out. I have come up with over 100 sentences now that I will put into the test network.

I have put together the GUI now, but as I am still working out some bugs, there is no functionality in it yet. In the meantime, I have an image of the second GUI mockup I made here. I am also working on sentences that the bot should spit out, divided into five or six different topics.

A new month again, and I am back in business after a short break. I have now completed the button and scrollbutton widgets, so the only thing left is the "widget" to show the network, as well as a bit of bugfixing. It looks like I will be able to actually put together my GUI this week.

I am done with the textbox now, except for a few bugs concerning newlines. I will not be able to work on the project next week since I am going out of town for a few days, and I also have other things to take care of.

I have begun implementing my GUI toolkit. So far I have done the main window and label widgets, and I am working on the textbox (multiline included). I expect to complete the textbox tomorrow, with editing and scrolling features.

I have done a first mockup of the GUI, it can be found here. I have also designed the widgets that I am going to create, and will begin implementing them next week.

I have spent most of the day trying out various graphical toolkits that use CLX, since I wanted to save some time by not having to write the code for widgets and things like that myself. However, I could not get any of them to work (except for one which was too buggy to use), so I will be writing the code myself after all. Hopefully it will not take longer than a couple of weeks.

I have started looking into using CLX for creating a GUI, to better visualize the networks. I will spend this week at least trying to create functions to display and alter the network nodes and parameters.

I have implemented creation of a uniform network with the number of nodes and link weight value as parameters. Tomorrow I will verify stability of uniform networks and do some other work as well.

I had another meeting with my supervisor today, and the basic functionality of my system is now considered stable. I was given a nice list of subgoals for the implementation part, and will start working on them next week. That is all for this week.

I have begun documenting the test networks in the report, and will continue doing that tomorrow as well. I am going away this weekend so I will not get any work done on Friday, but will post an update tomorrow.

October is here and things are looking bright again. Last week was not a productive one, since things did not work the way I wanted them to, so I lost some motivation. But I went back to the drawing board and started thinking some more about the problem, and I came up with a solution that I have implemented and tested today, beginning again with the smallest network of three nodes. The thing I had done wrong involved the resetting of nodes after they had fired, so now that part is working as intended, and I can now start documenting the tests I run on the two small example networks I have created. I will also try to update this page more often than I did last week, when I did not really have anything new to report.

I have spent most of the day adding error-handling to some calculations, since I got underflow errors while experimenting with the weights. Hopefully I can get some good results tomorrow that will be worth documenting.

I have begun experimenting with the weights and the parameters of the network. After getting a loop (the same node firing all the time), I realized that I have forgotten to add a random value to each calculated node. I will implement that code tomorrow and also continue with testing weight values and parameters.

Yesterday I had another meeting with my supervisor, and it looks like I am making good progress. I have added a function that checks if the activity level of a node is above a threshold value, and prints the node label if that is the case. Next week I will experiment with the weights in my example network and document the results in my report.

The UI is now done, and the neural network functions as well. I have added a new structure for the neural network itself instead of having it as a hash table. This means that the nodes, along with network parameters, are now encapsulated within this structure. I have also created a network with seven nodes, and I have run some tests on this network yielding results that look OK. I am going to set up a meeting with my supervisor so he can have a look as well.

I am done connecting the UI to the neural network functions, and tomorrow I will create a slightly larger network to test.

I have now converted all functions to using hash tables instead of lists where needed. I have also done quite a bit on a text user interface, which will facilitate testing and demonstration. Tomorrow I will continue with the UI and connect it to the functions that I have done so far.

Today I added a function for calculating the attention level, beta, of a node. I have also changed the neural network type from a list to a hash table. This means changing just about every function I have written so far, so I still have a bit left on that. Work on the project will continue on Monday.

I had another meeting with my supervisor today, and showed my work so far. I received some good feedback and also a reading tip, so I went to the library and borrowed "Introduction to the theory of neural computation" by Hertz, Krogh and Palmer. I have not done that much coding today, so I will not give a status report on that until tomorrow.

The activation level for every node in the network can now be calculated, and I have started on a function that lets you "touch" a node so it will get an activation level above the threshold. The function will then print out any nodes that gets activated as the activation spreads through the network.

Today I did not get everything done that I stated yesterday. I have set up the links and weights (including bias) for the three nodes in the network. I have made a print function for the node structure so that it is easier to see the values of the nodes. Finally, I have made a function to calculate the activity level of a node. Tomorrow I will continue and do the activity level calculation for the whole network, and with a bit of luck I might be able to test the network. I will make no promises though. I am learning Lisp as I go, so things take a bit longer than estimated.

I have started writing code for a simple recurrent network. In the current state I have hardcoded a small example with three nodes, and I have also created code for connecting these nodes and setting the weights on the links. Tomorrow I will implement the algorithm for calculating activation levels, and see if I can get my small network to work.

I am now finished with the specification, it can be found here (PDF).

Today I had a meeting with my supervisor. We went through what I should start with in the project and what other parts that need to be done. I have also gotten a bit further in the Lisp Primer. I finished the chapter about recursion, so I still have a bit to go, but I am now familiar with the syntax and some functions.

Tomorrow I will start writing the specification, and if there is time I will continue learning Lisp.

I've put up this temporary page for my MT, so I can post some updates while I work on a new version of my website.