Pacman Trainer

The final display with terminal debug output

Prerequisites

• Python 3 (so far, it works with Python 3, but if there are problems, try Python 2)

Environments Tested

Windows 10 WSL2 Ubuntu (Build 19042.1288)

• Git Bash
• Python 3.8.10 (Window WSL2 Ubuntu)
• Python 3.9.8 (See latest versions)
  • Python must be 3.9.x or earlier or PyTorch will not work

sudo apt update
sudo apt-get install python3-venv python3-tk
sudo apt-get install python3-tk

MacOS 11.6 (Build 20G165)

• Python 3.9.4

brew install python-tk

Quickstart

$ git clone git@github.com:masaok/pacman-agent.git
$ cd pacman-agent
$ python3 -m venv env
$ source env/Scripts/activate  # Windows Git Bash only
$ source env/bin/activate      # Mac / Linux only
$ pip install -r requirements.txt
$ python3 -u environment.py -h  # Show command-line help
$ python3 -u environment.py     # Run the app and show the GUI

$ deactivate  # exit virtual environment

VcXsrv (X server for Windows to run GUI stuff)

Download and install VcXsrv from here: https://sourceforge.net/projects/vcxsrv/

When running it (it's called XLaunch in Windows), make sure to check this box:

Also, add this to your .bash_profile and source it:

export DISPLAY=$(awk '/nameserver / {print $2; exit}' /etc/resolv.conf 2>/dev/null):0
export LIBGL_ALWAYS_INDIRECT=1

Verify that VcXsrv is running correctly by testing Xcalc:

Also, make this change in the code whereever torch.load is called:

https://stackoverflow.com/questions/56369030/runtimeerror-attempting-to-deserialize-object-on-a-cuda-device/62327502#62327502

Run the real thing:

python3 -u environment.py

Components

• PacmanAgent: Contains all of the logic for the agent controlling Pacman
• PacmanTrainer: Contains the logic for training Pacman using Pytorch for deep imitation learning.
• Environment: Contains the game logic and is largely structured in the same way as the Wumpus World / Blindbot's environment
  • A constructor, parameterized by the Pacman Trainer's array-of-strings maze representation, that initializes all of the maze variables, as follows:
    • The maze structure, where walls are located, and thus what legal moves are available in any position.
    • The sets and positions of all game actors (beginning with their provided starting positions), including:
      • Pacman
      • Ghosts
      • Pellets
  • Pacman's score (# of pellets eaten)
  • Determining game-ending conditions (viz., Pacman getting eaten by a ghost or eating every pellet)
  • Rendering the maze elements (in their basic array-of-strings format like in the Pacman Trainer) in terms of their graphical equivalents
  • Actors in the environment take turns acting, so the environment determines what happens on every "tick" or turn of a running game:
    • The PacmanAgent's chooseAction method is called with the current game state, and its action choice is enacted (if it's not a legal action, i.e., one that runs it into a wall, it does nothing that turn).
      • If Pacman eats a pellet, it is removed from the board and the score is incremented
      • If Pacman moves onto a ghost's tile, it dies and the game's over
    • All ghosts make an action choice governed by a coin-flip: 10% of the time, it will choose randomly, and the other 90%, it will take a step that brings it closer to Pacman. You can use / adapt the Pathfinder class I gave in the BlindBot package to help with pathfinding or just do something basic like looking at the Manhattan distance between a ghost and Pacman and then choosing the action that minimizes it.
• MazeUI: Given a Tk "root window", draw a Tk Canvas on it and all the blocks, characters, and items that make up a maze

Exercise

To view the accompanying exercise binding Pacman Trainer and Agent, see the EXERCISE.md instructions.