An AI was taught to play the most difficult video game in the world


What is the hardest video game you have ever played? If not QWOP So let me tell you well, I know that you don’t know how really difficult a game can be. The deceptively simple racing game is so challenging to master that even a AI trained through machine learning still only collected a top 10 score instead of breaking the record.

If you have never juice QWOP prior to, you owe it to yourself give it a try and see if you can get your sprinter off the starting line. Developed by Bennett foddy in 2008, QWOP was inspired by an 80s arcade game called Track and field that requires players to mindlessly pounding buttons to win a race. QWOP takes a different approach and instead makes players use four keys to control individual movements of a runner thighs and calves-to runner who behaves like a flexible rag doll and is subject toworld physics, including the effects of gravity. It may sound simple, but mastering the timing and cadence of keystrokes required to get the sprinter awkwardly forward can be incredibly frustrating.

Wesley Liao was curious to know how well a tool like AI, which has been trained to do things like realistically animated old pictures of deceased loved ones, I would play QWOP. After first creating a Javascript adapter that would allow an AI tool to actually play and interact with the game, Liao’s first attempt at machine learning simply had the AI ​​play the game on its own and learn what actions resulted in results. positive (sprinter advancing and increasing their speed) and which resulted in negative results (the sprinter’s torso flex too close to the ground.) Through this approach, the AI ​​learned a “knee scraping” technique that would make it pass successfullymeters from the finish line, but not at record speeds.

Liao’s next attempt to train an AI model involved recording gameplay videos of them trying to be successful in the game, including using longer strides that are crucial to increasing speed and crossing the finish line with decent time. . The approach was slightly more successful, but the AI ​​was unable to master a special technique used by advanced QWOP players that involve an upward and forward leg movement to generate additional momentum.

Finally, Liao approached a veteran player known as Kurodo (@cld_el on Twitter), one of the best QWOP sprinters in the world, who recorded 50 videos of themselves playing at an expert level. But even with access to the best possible game techniques, Liao found that the best results came from a machine learning training regimen that involved 25 hours of AI play alone, 15 hours of learning from the data obtained from the Kurodo expert races and another 25 hours of auto-play.

But even with all that effort, the QWOP-playing the top 100 of AI-The result of the subway board made him cross the finish line in 1 minute and 8 seconds—a top 10 Finalize. According to Speedrun.com, the current world record for 100-meter running is just 48 seconds, set just a month ago. Liao is confident of having more training and a different reward system (how the AI ​​learns that it has done something correctly), to establish a QWOP The world record could eventually happen, although since it is a computer that plays the game, the record may never be officially recognized.

.

Source link