#Leaderboard

#Bounties

Leaderboard

Bounties

The second will be 100 points, which I will award when any of the following occurs:

• an answer is posted based on neural networks and gets a reasonable score. (It doesn't need to compete with the top answers, just do reasonably well.)

• an answer is posted that takes a different method from the existing answers, in my subjective opinion. All the current top answers use variations on a Markov model (i.e. store substrings of length n and guess the most common next letter); I'm looking for an answer that does something significantly different from this. As above, this should score reasonably well but need not compete with the top answers.

• an answer is posted that (a) uses a sophisticated method (in my subjective opinion), and (b) has a good explanation of that method, where "good explanation" means I could implement it myself without reading the code. This bounty could be awarded to an existing answer, if one is updated with a good explanation.

Once the 100 point bounty is claimed I will offer a 200 point one, updating the requirements accordingly.

The second, 100 points, was also awarded to A. Rex, for the same answer, because they added a very good explanation to their existing answer.

The next bounty, 200 points, will be awarded to either

• A competitive answer that uses a new technique. (This will be based on my subjective judgment since it's my rep that goes into the bounty, but you can trust me to be fair. Note that your answer needs to contain sufficient explanation for me to understand how it works!) Such an answer needn't take the top score, it just needs to do reasonably well compared to existing answers. I'm particularly keen to see solutions based on recurrent neural networks, but I'll award the bounty to anything that seems different enough from the Markov models that dominate the current top scores.

Or:

• Anyone else who beats A. Rex's top score (currently 444444), using any method.

Once the 200 point bounty is claimed I will most likely offer a 400 point one, updating the requirements accordingly.

There are two versions of the file, and you may use either of them in an answer:

• whale2.txt (linked above) - in this version the text is not wrapped, so newlines appear only at the end of paragraphs. This file is recommended for new answers.

• whale.txt - in this version the text is wrapped to a width of 74 characters, so you have to predict the end of each line as well as predicting the text. This makes the challenge more fiddly and is included for mostly historical reasons.

Both files are the same size, 1215236 bytes.

• Your submission itself. (The code, plus any data files it uses - these can be links if they're large.)
• An explanation of how your code works. Please explain the I/O method as well as how it predicts the next character.
• The code you used to evaluate your score. (If this is identical to a previous answer you can just link to it.)
• Any code you used to generate your submission, along with an explanation of that code. This includes code that you used to optimise parameters, generate data files etc. (This doesn't count towards your byte count but should be included in your answer.)

The first one will be 50 points, to be awarded to the best-scoring answer posted during the bounty period.

The second will be 100 points which I will award to the first answer that takes a different approach from the existing ones, according to my subjective judgement. (I'll create the bounty at the time of awarding it.) I'm particularly keen to see an approach based on neural networks, but I'll also award it to other techniques that are sufficiently different from the existing Markov model based approaches. Answers don't need to be among the top scoring to compete for this bounty, but ideally they should beat the median score. It is possible for one answer to win both bounties.

For historical reasons, there are two versions of the file, and you may use either of them in an answer. In whale2.txt (linked above) the text is not wrapped, so newlines appear only at the end of paragraphs. In the original whale.txt the text is wrapped to a width of 74 characters, so you have to predict the end of each line as well as predicting the text. This makes the challenge more fiddly, so whale2.txt is recommended for new answers. Both files are the same size, 1215236 bytes.

• Your submission itself. (The code, plus any data files it uses - these can be links if they're large.)
• An explanation of how your code works. Please explain the I/O method as well as how it predicts the next character. The explanation of your algorithm is important, and good explanations will earn bounties from me.
• The code you used to evaluate your score. (If this is identical to a previous answer you can just link to it.)
• Any code you used to generate your submission, along with an explanation of that code. This includes code that you used to optimise parameters, generate data files etc. (This doesn't count towards your byte count but should be included in your answer.)

The first one, 50 points, was awarded to A. Rex for the best-scoring answer at the time.

The second will be 100 points, which I will award when any of the following occurs:

• an answer is posted based on neural networks and gets a reasonable score. (It doesn't need to compete with the top answers, just do reasonably well.)

• an answer is posted that takes a different method from the existing answers, in my subjective opinion. All the current top answers use variations on a Markov model (i.e. store substrings of length n and guess the most common next letter); I'm looking for an answer that does something significantly different from this. As above, this should score reasonably well but need not compete with the top answers.

• an answer is posted that (a) uses a sophisticated method (in my subjective opinion), and (b) has a good explanation of that method, where "good explanation" means I could implement it myself without reading the code. This bounty could be awarded to an existing answer, if one is updated with a good explanation.

Once the 100 point bounty is claimed I will offer a 200 point one, updating the requirements accordingly.