User:Matthew Elliott/Diary

From seed
Jump to: navigation, search

February 25, 2016: MindReader is up!

The new scoreboard is up for the MindReader site.

I am almost done building a player profile page for the users as well as an "about" page which describes the website.

February 3, 2016: MindReader is up!

The time version of MindReader is now available just in time for the ITA conference.

Next steps are to add the following:

  • A consecutive scores board
  • A hour, day, month, ever score board
  • Build Facebook page

January 6, 2016: MindReader is up!

The MindReader game is up and can be played at:

--Yoavfreund 09:01, 7 January 2016 (PST) Excellent! What remains in terms of the GUI is to collect a name for the user and to create a real leader-board.

The games dataset and details about the website may be accessed on Google's Developer Console [1]. Follow link [1] and sign in using the email address: The Google Drive [2] associated with this gmail account contains all the source code and files for this project.

--Yoavfreund 08:56, 7 January 2016 (PST) Looks great! but where do I find the log files?

The algorithm used for this MindReader game is an exact replica of the default algorithm used in Prof. Schapire's C code. This is a hedging algorithm with 2 experts. To verify that the algorithm is correct I ran both my MindReader and Schapire's MindReader using the same string of inputs. At each step I had both algorithms output the updated weights of the experts and had them output the Bernoulli probability used to produce the computer next guess, F(r) . To make sure the algorithms were the same, I made the computers next guess deterministic. This was done by rounding the Bernoulli probability, F(r), to the nearest integer (1 or 0) instead of using the Bernoulli distribution.

My MindReader[3], and Schapire's MindReader[4] produced the exact same results at every step. This can be verified by comparing link [3] and [4]. Additionally, I checked that the results were the same when each expert was run individually. If further verification is required, I could run both algorithms multiple times with the same input sequence and then check to see that the total loss of each algorithm follows the same probability distribution.

--Yoavfreund 08:56, 7 January 2016 (PST) An alternative to running many times is to have both algorithm output F(r) and compare those (the comparison should be approximate, because the roundoff error is likely to be different in the two implementions.

--Yoavfreund 09:01, 7 January 2016 (PST) I think the next big step would be to collect enough data that you can compare different sets of experts on it.

September 27, 2016: Summer Work Summary

The main addition is adding a Facebook API. Time was spent looking into Google's abilities of making asynchronous calls to the backend, however, this was shown as not being a viable option. A system that achieves a relatively similar level of security was setup using post request though.

Next steps are to add the following:

  • Restructuring backend database
  • Shifting high score to point based mechanism
  • Analyzing data on an ipython server.