http://www.robotworldnews.com/100389.php
This is the first time I have ever heard of an AI being designed for a video game that wasn't designed by the original programmers. I wonder if any of these techniques could be used to make better TAS tools.
Tetris is quite popular for AI, at least I thought so. It's a fairly simple game at the core.
But:
AI agents who learned with the most successful policy had an average score of 8186 points over 50 games. The average score of five humans, who each played 10 games, was 8064. In the non-learning experiment when agents used random policies, the average score was just 676. Other policies produced scores that fell within this range.
I can usually score a decent average of 70,000, and I am by no means good...
Wouldn't it have made more sense to have, say, 10-15 humans and take the top 5 to get a better guess?
Joined: 11/18/2006
Posts: 2426
Location: Back where I belong
Xkeeper wrote:
Wouldn't it have made more sense to have, say, 10-15 humans and take the top 5 to get a better guess?
I may have made more sense, but that's not how scientific studies work (and I believe this was supposed to be a publishable journal article, not just a "look what we did" experiment). Imagine if they did the same thing with new cancer drugs...
Tetris is quite popular for AI, at least I thought so. It's a fairly simple game at the core.
But:
AI agents who learned with the most successful policy had an average score of 8186 points over 50 games. The average score of five humans, who each played 10 games, was 8064. In the non-learning experiment when agents used random policies, the average score was just 676. Other policies produced scores that fell within this range.
I can usually score a decent average of 70,000, and I am by no means good...
Wouldn't it have made more sense to have, say, 10-15 humans and take the top 5 to get a better guess?
I seem to remember me scoring around 100,000 points when I first played ms. pacman. I'm sure the people they used for their research do not play videogames.
I thought that an average score of the first level being beaten would be a score of around 8,000? maybe I'm wrong.
http://www.pacmangame.info/ms_pacman.html
Well, if this game scores the same way the original does, I got a score of 11,300 for the first level trying my best with one life.
Super Mario Bros. console speedrunner
- Andrew Gardikis
Wouldn't it have made more sense to have, say, 10-15 humans and take the top 5 to get a better guess?
I may have made more sense, but that's not how scientific studies work (and I believe this was supposed to be a publishable journal article, not just a "look what we did" experiment). Imagine if they did the same thing with new cancer drugs...
Yes, but read the article again:
...with the most successful policy...
A rough equivalent of this would probably be taking only the best n players average.
I am not saying that the results should be skewed for no reason.
