'AI playing Hudson\'s Adventure Island (NES) using NEAT - try 1 - max fitness 7180'

Posted Sep 14, 2022

'So this is my stream of my PC running a Bizhawk emulator with Hudson\'s Adventure Island (NES) and some AI logic which is still in development. It is very crude right now: not only does it not recognize the ground so it can\'t finish the fourth level, the resolution is also bad and the network/map still needs a lot of tweaking.  But while I work out the details I thought it would be interesting to have it run and see how good (or bad) this goes and if it can get as far as the pits.   The AI is written in Lua and is basically a modified version of MarI/O: https://github.com/pakoito/MarI-O  All credits go to SethBling ( http://sethbling.com/ & https://www.youtube.com/channel/UC8aG3LDTDwNR1UQhSn9uVrw ). While I have a very very basic understanding of neural networks and the associated algorithms, I know shit all about implementing them. This gave me a chance to headstart into doing things without having to worry (yet) about some of the details.  The AI is basically an implementation of the NEAT or \"NeuroEvolution of Augmenting Topologies\" method, described in this paper: http://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf It is basically a neural network which employs a self-evolving topology (the way neurons are connected) instead of a more traditional fixed topology. It uses a genetic alghorithm where not only the weights within the network, but the topology itself is part of the species. Some random specifics:  - Fitness is (lazily) calculated based on the Y position of the sprite on the left showing progress through the level - Input is based on the grid you see overlayed. Values are based on the X/Y positions of \"enemies\" (value -1, black) or the side of the road (value 1, white), as read from memory - Timeout is based on actually not progressing. Failure is based on fuel = 0. Actually finishing a level hasn\'t been implemented yet - Software is Bizhawk, and its awesome. It lets you read and analyze RAM, vital to get the necessary variables for calculating the input map' 

See also: iifym , 26 , 1rm , Training � , merengue , fat loss plan , workout motivation , journey , no equipment , lan

comments