Project 4: Ghostbusters

(Thanks to John DeNero and Dan Klein!)

Enough of defense,
Onto enemy terrain.
Capture all their food!


The course contest involves a multi-player capture-the-flag variant of Pacman, where agents control both Pacman and ghosts in coordinated team-based strategies. Your team will try to eat the food on the far side of the map, while defending the food on your home side.

Key files to read: The main file that runs games locally. This file also describes the new capture the flag GameState type and rules. Specification and helper methods for capture agents. Example code that defines two very basic reflex agents, to help you get started. This is where you define your own agents for inclusion in the nightly tournament. (This is the only file that you submit.)
Supporting files (do not modify): The logic behind how the Pacman world works. This file describes several supporting types like AgentState, Agent, Direction, and Grid. Useful data structures for implementing search algorithms. Computes shortest paths between all maze positions. Graphics for Pacman Support for Pacman graphics ASCII graphics for Pacman Keyboard interfaces to control Pacman Code for reading layout files and storing their contents

Academic Dishonesty: While we won't grade contests, we still expect you not to falsely represent your work. Please don't let us down.

Rules of Pacman Capture the Flag

Layout: The Pacman map is now divided into two halves: blue (right) and red (left). Red agents (which all have even indices) must defend the red food while trying to eat the blue food. When on the red side, a red agent is a ghost. When crossing into enemy territory, the agent becomes a Pacman.

Scoring: When a Pacman eats a food dot, the food is permanently removed and one point is scored for that Pacman's team. Red team scores are positive, while Blue team scores are negative.

Eating Pacman: When a Pacman is eaten by an opposing ghost, the Pacman returns to its starting position (as a ghost). No points are awarded for eating an opponent.

Power capsules: If Pacman eats a power capsule, agents on the opposing team become "scared" for the next 40 moves, or until they are eaten and respawn, whichever comes sooner. Agents that are "scared" are susceptible while in the form of ghosts (i.e. while on their own team's side) to being eaten by Pacman. Specifically, if Pacman collides with a "scared" ghost, Pacman is unaffected and the ghost respawns at its starting position (no longer in the &lquot;scared&rquot; state).

Observations: Agents can only observe an opponent's configuration (position and direction) if they or their teammate is within 5 squares (Manhattan distance). In addition, an agent always gets a noisy distance reading for each agent on the board, which can be used to approximately locate unobserved opponents.

Winning: A game ends when one team eats all but two of the opponents' dots. Games are also limited to 1200 agent moves (300 moves per each of the four agents). If this move limit is reached, whichever team has eaten the most food wins. If the score is zero (i.e., tied) this is recorded as a tie game.

Computation Time: We will run your submissions on the WashU cluster (like all your assignments). Each agent has 1 second to return each action. Each move which does not return within one second will incur a warning. After three warnings, or any single move taking more than 3 seconds, the game is forfeited. There will be an initial start-up allowance of 15 seconds (use the registerInitialState function). If you agent times out or otherwise throws an exception, an error message will be present in the log files, which you can download from the results page (see below).

Getting Started

By default, you can run a game with the simple baselineTeam that the staff has provided:


A wealth of options are available to you:

python --help

There are four slots for agents, where agents 0 and 2 are always on the red team, and 1 and 3 are on the blue team. Agents are created by agent factories (one for Red, one for Blue). See the section on designing agents for a description of the agents invoked above.

The only team that we provide is the baselineTeam. It is chosen by default as both the red and blue team, but as an example of how to choose teams:

python -r baselineTeam -b baselineTeam

which specifies that the red team -r and the blue team -b are both created from

To control one of the four agents with the keyboard, pass the appropriate option:

python --keys0

The arrow keys control your character, which will change from ghost to Pacman when crossing the center line.


By default, all games are run on the defaultcapture layout. To test your agent on other layouts, use the -l option. In particular, you can generate random layouts by specifying RANDOM[seed]. For example, -l RANDOM13 will use a map randomly generated with seed 13.

All layouts are symmetric, and the team that moves first is randomly chosen.

Designing Agents

Unlike project 2, an agent now has the more complex job of trading off offense versus defense and effectively functioning as both a ghost and a Pacman in a team setting. Furthermore, the limited information provided to your agent will likely necessitate some probabilistic tracking (like project 4). Finally, the added time limit of computation introduces new challenges.

Baseline Team: To kickstart your agent design, we have provided you with a team of two baseline agents, defined in They are both quite bad. The OffensiveReflexAgent moves toward the closest food on the opposing side. The DefensiveReflexAgent wanders around on its own side and tries to chase down invaders it happens to see.

Interface: The GameState in should look familiar, but contains new methods like getRedFood, which gets a grid of food on the red side (note that the grid is the size of the board, but is only true for cells on the red side with food). Also, note that you can list a team's indices with getRedTeamIndices, or test membership with isOnRedTeam.

Finally, you can access the list of noisy distance observations via getAgentDistances. These distances are within 6 of the truth, and the noise is chosen uniformly at random from the range [-6, 6] (e.g., if the true distance is 6, then each of {0, 1, ..., 12} is chosen with probability 1/13). You can get the likelihood of a noisy reading using getDistanceProb.

Distance Calculation: To facilitate agent development, we provide code in to supply shortest path maze distances.

To get started designing your own agent, we recommend subclassing the CaptureAgent class. This provides access to several convenience methods. Some useful methods are:

  def getFood(self, gameState):
    Returns the food you're meant to eat. This is in the form
    of a matrix where m[x][y]=true if there is food you can
    eat (based on your team) in that square.

  def getFoodYouAreDefending(self, gameState):
    Returns the food you're meant to protect (i.e., that your
    opponent is supposed to eat). This is in the form of a
    matrix where m[x][y]=true if there is food at (x,y) that
    your opponent can eat.

  def getOpponents(self, gameState):
    Returns agent indices of your opponents. This is the list
    of the numbers of the agents (e.g., red might be "1,3")

  def getTeam(self, gameState):
    Returns agent indices of your team. This is the list of
    the numbers of the agents (e.g., red might be "1,3")

  def getScore(self, gameState):
    Returns how much you are beating the other team by in the
    form of a number that is the difference between your score
    and the opponents score. This number is negative if you're

  def getMazeDistance(self, pos1, pos2):
    Returns the distance between two points; These are calculated using the provided
    distancer object.

    If distancer.getMazeDistances() has been called, then maze distances are available.
    Otherwise, this just returns Manhattan distance.

  def getPreviousObservation(self):
    Returns the GameState object corresponding to the last
    state this agent saw (the observed state of the game last
    time this agent moved - this may not include all of your
    opponent's agent locations exactly).

  def getCurrentObservation(self):
    Returns the GameState object corresponding this agent's
    current observation (the observed state of the game - this
    may not include all of your opponent's agent locations

  def debugDraw(self, cells, color, clear=False):
    Draws a colored box on each of the cells you specify. If clear is True,
    will clear all old drawings before drawing on the specified cells.
    This is useful for debugging the locations that your code works with.

    color: list of RGB values between 0 and 1 (i.e. [1,0,0] for red)
    cells: list of game positions to draw on  (i.e. [(20,5), (3,22)])

Restrictions: You are free to design any agent you want. However, you will need to respect the provided APIs if you want to participate in the tournaments. Agents which compute during the opponent's turn will be disqualified. In particular, any form of multi-threading is disallowed, because we have found it very hard to ensure that no computation takes place on the opponent's turn.

Warning: If one of your agents produces any stdout/stderr output during its games in the nightly tournaments, that output will be included in the contest results posted on the website. Additionally, in some cases a stack trace may be shown among this output in the event that one of your agents throws an exception. You should design your code in such a way that this does not expose any information that you wish to keep confidential.

Official Tournaments

The actual competitions will be run using nightly automated tournaments, with the final tournament deciding the final contest outcome. Tournaments are run everyday at approximately 11pm and include all teams that have been submitted (either earlier in the day or on a previous day) as of the start of the tournament. Currently, each team plays every other team 3 times (to reduce randomness), but this may change later in the semester. The layouts used in the tournament will be drawn from both the default layouts included in the repos as well as randomly generated layouts each night.

Submission Instructions

To enter into the tournament, your team must be defined in a file whose filename matches your codename.txt file exactly (including capital letters etc.). For example, if your codename.txt specifies your team-name as Superman then you must add a file into the repository. To do so, first copy the file to and then call svn add and svn commit -m "added my awesome agent". Due to the way the tournaments are run, your code must not rely on any additional files that we have not provided (The submission system may allow you to submit additional files, but the contest framework will not include them when your code is run in the tournament). You may not modify the code we provide.

As usual, you should also include a partners.txt file. Different from previous assignments, this time your codename must be only ASCII characters and cannot contain spaces etc.

Contest Details

Teams: You may work in teams of up to 5 people.

Scoring: Rankings are determined according to the points achieved in individual games. Currently any commit leads to a recomputation of the leaderboard. This might change if the computational burden becomes too high (in which case we will switch to nightly competitions).

This project is out of 20 points total. You will get fourteen points for beating the BASELINE program. You can earn extra points for winning certain matches. You get 2 points for beating the other benchmarks, Bender, Jarvis, Sonny and Random. You can also gain extra points for having the best agent on particular days.

If you have one of the top two agents on 11/24 (Tuesday before Thanksgiving) at 2:30pm, then you gain an extra point. If you have one of the top three agents on 12/3 (our last lecture) at 2:30pm, you also gain an extra point. Finally, if you have one of the top four agents at the end of the class, 2:30 on 12/16, then you gain one additional point.


Thanks to Barak Michener and Ed Karuna for providing improved graphics and debugging help.