This is an implementation of a Ms. Pacman game simulation and Q-learning AI agent in C++.
The Q-learning function is approximated using a multilayer feed-forward neural network.
It was programmed entirely from scratch, following guidelines from the following papers:
- L. Bom et al, Reinforcement Learning to Train Ms. Pac-Man Using Higher-order Action-relative Inputs, University of Groningen, Groningen, The Netherlands
- D. Ballard et al, Evolving Multimodal Behavior Through Modular Multiobjective Neuroevolution, University of Texas, Austin, USA
- D. B. Ortega, Machine Learning Applied to Pac-Man, Barcelona School of Informatics, UPC, Barcelona, Spain
- M. H van der Ree and M. A Wiering, Reinforcement Learning in the Game of Othello: Learning Against a Fixed Opponent and Learning from Self-Play, University of Groningen, Groningen, The Netherlands
And with the help and supervision of professor Marco A. Wiering from the University of Groningen.
Disclaimer: Only tested on Unix platforms
git clone https://github.com/albertsgrc/q-mspacman.gitcd q-mspacman/srcmake
In case you want to execute the scripts provided inside the src/scripts folder,
you should also install coffeescript via npm.
Node, together with npm, can be downloaded and installed from the following link:
https://nodejs.org/en/download/
Once node is installed, run npm install -g coffee-script (may need privileges)
in order to install the coffee-script interpreter. A script named script.coffee
can be run with the command coffee script.coffee.
The generated executable is called pacman. It accepts several program arguments,
which can be specified as in the following example command:
./pacman argname1=value1 argname2=value2 (etc.)
Every argument has a default value. For the full list of arguments, its meanings
and default values refer to the source file arguments.hh.
In order to train a neural network with the default maze originalClassic.lay,
one can simply run ./pacman. The resulting neural network is saved in the file
data/neural-networks/nn<id>.txt. The id is generated automatically.
Now, in order to visualize the reinforcement learning agent game play, one can
run ./pacman pacman_ai_agent=nn<id> plays=1. Here <id> is the id of the
neural network that was generated, which can be obtained from the corresponding
file name. To ease the process, one can use last-nn.coffee script,
which outputs the name of the last generated neural network. The command would look like:
./pacman pacman_ai_agent=$(coffee scripts/last-nn.coffee) plays=1.
The best neural network that we obtained during our experiments and testing of the program is included in data/neural-networks/nnbest.txt. Its gameplay can be visualized with the command ./pacman pacman_ai_agent=nnbest plays=1.
Note that with the default arguments the training process takes a long time, because a high number of hidden layers and neurons is used in order to obtain better performance. Still, by using less neurons and training steps good performance (> 60% win ratio) can still be achieved. One example of such parameters would be:
pacman n_hidden_layers=1 n_hidden_neurons=150 plays=7500 learning_rate=0.0004 activation_function=sigmoid