5 ms·

Related: there are various bots for HQ Trivia that do similar things that have come up on HN before. Here's one with a bit of a writeup: https://medium.com/@to

by splonk 8y ago

Related: there are various bots for HQ Trivia that do similar things that have come up on HN before. Here's one with a bit of a writeup: https://medium.com/@tobymellor/hq-trivia-using-bots-to-win-money-from-online-game-shows-ce2a1b11828b https://medium.com/@tobymellor/hq-trivia-using-bots-to-win-m... My general impression is that being able to screenshot and use real computing power is a fair bit easier than using OCR on a Raspberry Pi, but question parsing and querying Google is a lot harder than using a static corpus of questions and answers.

applecrazy 8y ago
I've built a similar proof-of-concept HQ Trivia algorithm before[0], and from my simple tests, in most cases isn't any additional work needed, just naively send the OCRed question to Google. Of course, some questions don't work using this naive approach, which is why a similar HQ Trivia bot used riffed on this concept with unique (and clever) heuristics[1]. And yes, it is easier to work with a screenshot than a grainy image of text produced by a camera. [0]: https://applecrazy.github.io/blog/posts/protect-trivia-from-bots/ https://applecrazy.github.io/blog/posts/protect-trivia-from-... [1]: https://hackernoon.com/i-hacked-hq-trivia-but-heres-how-they-can-stop-me-68750ed16365 https://hackernoon.com/i-hacked-hq-trivia-but-heres-how-they... Edit: my HQ Trivia system was a simple 1-hour weekend project, so it is by no means perfect. Also, to prevent my Github from being taken down, none of the code is available.
bdod6 8y ago
Mux wrote a blog post about this too: https://mux.com/blog/how-we-used-machine-learning-to-predict-hq-trivia-answers/ https://mux.com/blog/how-we-used-machine-learning-to-predict... Using machine learning to parse text helps, but it's still not a perfect solution.
kregasaurusrex 8y ago
I'd previously done OCR work with Twitch streams, but the reliability of the Tesseract libraries for custom fonts had stopped me from going any further with it. Since HQ uses a standardized font library, I'd imagine character extraction would be more successful when recording from screenshots because it would likely be closer to the library's training datasets.
- lathiat 8y ago
  Counting donations and subs? I’d thought about doing the same.
- driverdan 8y ago
  You can train Tesseract on other fonts.
pascoej 8y ago
Even with a good bot (displays info instead of likely answer), HQ stops you from withdrawing after you win too many games, no fun :( You can directly connect to their websocket and get the text of the answers faster than they are displayed on the app.

by splonk 8y ago

applecrazy 8y ago

I've built a similar proof-of-concept HQ Trivia algorithm before[0], and from my simple tests, in most cases isn't any additional work needed, just naively send the OCRed question to Google. Of course, some questions don't work using this naive approach, which is why a similar HQ Trivia bot used riffed on this concept with unique (and clever) heuristics[1]. And yes, it is easier to work with a screenshot than a grainy image of text produced by a camera. [0]: https://applecrazy.github.io/blog/posts/protect-trivia-from-bots/ https://applecrazy.github.io/blog/posts/protect-trivia-from-... [1]: https://hackernoon.com/i-hacked-hq-trivia-but-heres-how-they-can-stop-me-68750ed16365 https://hackernoon.com/i-hacked-hq-trivia-but-heres-how-they... Edit: my HQ Trivia system was a simple 1-hour weekend project, so it is by no means perfect. Also, to prevent my Github from being taken down, none of the code is available.

bdod6 8y ago

Mux wrote a blog post about this too: https://mux.com/blog/how-we-used-machine-learning-to-predict-hq-trivia-answers/ https://mux.com/blog/how-we-used-machine-learning-to-predict... Using machine learning to parse text helps, but it's still not a perfect solution.

kregasaurusrex 8y ago

I'd previously done OCR work with Twitch streams, but the reliability of the Tesseract libraries for custom fonts had stopped me from going any further with it. Since HQ uses a standardized font library, I'd imagine character extraction would be more successful when recording from screenshots because it would likely be closer to the library's training datasets.

lathiat 8y ago
Counting donations and subs? I’d thought about doing the same.
driverdan 8y ago
You can train Tesseract on other fonts.

pascoej 8y ago

Even with a good bot (displays info instead of likely answer), HQ stops you from withdrawing after you win too many games, no fun :( You can directly connect to their websocket and get the text of the answers faster than they are displayed on the app.