A software engineer has trained an artificial intelligence to play Pokémon Red

Engineer Peter Whidden has developed an artificial intelligence that learned how to play the classic 1996 game Pokémon Red using reinforcement learning. The AI has spent more than 50,000 hours in the virtual world of Pokémon over the years.

Here's What We Know

Whidden posted a one-minute video on YouTube demonstrating how the AI works in the game. The developer also published the code and instructions he used on GitHub so that other users can create their own virtual players based on his algorithms.

The reinforcement model incentivises the AI to increase the level of Pokémon on a team, explore new locations, win battles and defeat stadium leaders. Sometimes these goals diverge from game progression, leading to amusing AI behaviour.

For example, the AI may get "stuck" in one place while admiring its surroundings, or experience "trauma" when accidentally losing a Pokémon. Such sub-optimal but endearing AI play makes the audience sympathetic.

Initially, the AI couldn't even get past the starting locations because it couldn't interpret the text clues in the game. Whidden made changes to the code and algorithms to help the virtual player progress further. This enabled the AI to reach the first caves outside of the initial city.

According to Whidden, this approach allows for an interesting explanation of the workings of AI algorithms using the example of a popular game. Reinforcement learning has previously been used to create algorithms that play chess, Go and other games. But a project based on Pokémon Red has attracted particular attention due to its use of favourite characters as a demonstration of complex AI concepts.

Source: TechCrunch