Homework Assignment #2: Multi-Agent Pacman Solution

Acknowledgements: This project is based on Berkeley’s CS188 EdX course project. It is a modified and

augmented version of “Project 2: Multi-Agent Pacman” available at http://ai.berkeley.edu/multiagent.html.

For this part of the project, you will design agents for the classic version of Pacman, including ghosts. Along the way, you will implement both minimax and expectimax search and try your hand at evaluation function design.

    • Files you will edit or create:

– multiAgents.py - Suitably augment this with your implementation of the search agents de-scribed in tasks 1 to 3 in this handout.

    • Files that are useful to your implementations:

– pacman.py - Runs Pacman games. This file also describes a Pacman GameState, which you will use extensively in your implementations.

– game.py - The logic behind how the Pacman world works. This file describes several support-ing types like AgentState, Agent, Direction, and Grid.

– util.py - Useful data structures for implementing search algorithms.

    • Files that are unlikely to be of help to your implementations, but are required to run the project.

– graphicsDisplay.py - Graphics for Pacman.

– graphicsUtils.py - Support for Pacman graphics.

– textDisplay.py - ASCII graphics for Pacman.

– ghostAgents.py - Agents to control ghosts

– keyboardAgents.py - Keyboard interfaces to control Pacman.

– layout.py - Code for reading layout files and storing their contents.

– autograder.py - Project autograder.

– testParser.py - Parses autograder test and solution files.

– testClasses.py - General autograding test classes.

– multiagentTestClasses.py - Project specific autograding test classes.

– grading.py - File that specifies how much each question in autograder is worth.

– projectParams.py - Project parameters, facilitates nicer output during autograding.

– pacmanAgents.py - Classes for specifying ghosts’ behaviour.


The Pacman Game

Pacman is a video game first developed in the 1980s. A basic description of the game can be found at https://en.wikipedia.org/wiki/Pac-Man. In order to run your agents in a game of Pacman, and to evaluate your agents with the supplied test code, you will be using the command line. To familiarize

Assignment 2, University of Toronto, CSC384 - Intro to AI,  
3

yourself with running this game from the command line, try playing a game of Pacman yourself by typing the following command:

python pacman.py

To run Pacman with a game agent use the -p command. Run Pacman as a GreedyAgent:

python pacman.py -p GreedyAgent

You can run Pacman on different maps using the -l command:

python pacman.py -p GreedyAgent -l testClassic

Important: If you use the command in an environment with no graphical interface (e.g. when you ssh to teach.cs, formerly CDF), you must use the flag --no-graphics, which suppresses graphical output.

The following commands are available to you when running Pacman. They are also accessible by running python pacman.py --help.

    • -q Allows you to select a game agent for controlling pacman, e.g., -p GreedyAgent. By the end of this project, you will have implemented four more agents, listed.

– ReflexAgent

– MinimaxAgent

– AlphaBetaAgent

– ExpectimaxAgent

    • -l Allows you to select a map for playing Pacman. There are 9 playable maps, listed.

– minimaxClassic

– trappedClassic

– testClassic

– smallClassic

– capsuleClassic

– openClassic

– contestClassic

– mediumClassic

– originalClassic

    • -a Allows you to specify agent specific arguments. For instance, for any agent that is a subclass of MultiAgentSearchAgent, you can specify the depth that you limit your search tree by typing -a depth=3

    • -n Allows you to specify the amount of games that are played consecutively by Pacman, e.g., -n 100 will cause Pacman to play 100 consecutive games.
Assignment 2, University of Toronto, CSC384 - Intro to AI,  
4

    • -k Allows you to specify how many ghosts will appear on the map. For instance, to have 3 ghosts chase Pacman on the contestClassic map, you can type -l contestClassic -k 3
Note: There is a max number of ghosts that can be initialized on each map, if the number specified exceeds this number, you will only see the max amount of ghosts.

    • -g Allows you to specify whether the ghost will be a RandomGhost (which is the default) or a DirectionalGhost that chases Pacman on the map. For instance, to have DirectionalGhost charac-ters type -g DirectionalGhost

    • --no-graphics Allows you to run Pacman with no graphics.

    • --frameTime Specifies frame time for each frame in the Pacman visualizer (e.g., --frameTime 0). An example of a command you might want to run is:

python pacman.py -p GreedyAgent -l contestClassic -n 100 -k 2 -g DirectionalGhost -q

This will run a GreedyAgent over 100 cases on the contestClassic level with 2 DirectionalGhost characters, while supressing the visual output.

Important Note: To run the autograder for all questions, run the following command:

python autograder.py

Note that the provided starter code also contains testcases. The autograder will run all tests.
Assignment 2, University of Toronto, CSC384 - Intro to AI,  
5

Question 1 : Reflex Agent (8 points)

Don’t spend too much time on this question, as the meat of the project lies ahead.

Improve the ReflexAgent in multiAgents.py to play respectably. The provided reflex agent code pro-vides some helpful examples of methods that query the GameState for information. A capable reflex agent will have to consider both food locations and ghost locations to perform well. Your agent should easily and reliably clear the testClassic layout:

python pacman.py -p ReflexAgent -l testClassic

Try out your reflex agent on the default mediumClassic layout with one ghost or two (and animation off to speed up the display):

python pacman.py --frameTime 0 -p ReflexAgent -k 1 python pacman.py --frameTime 0 -p ReflexAgent -k 2

How does your agent fare? It will likely often die with 2 ghosts on the default board, unless your evaluation function is quite good.

Note: As features, try the reciprocal of important values (such as distance to food) rather than just the values themselves.

Note: The evaluation function you’re writing is evaluating state-action pairs (i.e., how good is it to perform this action in this state); in later parts of the project, you’ll be evaluating states (i.e., how good is it to be in this state).

Options: Default ghosts are random; you can also play for fun with slightly smarter directional ghosts using -g DirectionalGhost. If the randomness is preventing you from telling whether your agent is improv-ing, you can use -f to run with a fixed random seed (same random choices every game). You can also play multiple games in a row with -n. Turn off graphics with -q to run lots of games quickly.

To test with the autograder, you may run:

python autograder.py -q q1



Question 2: Minimax (10 points)

Here you will implement a minimax search agent for the game of Pacman. The function getAction in the MinimaxAgent class can be found in multiAgents.py. Your minimax search must work with any number of ghosts. Your search tree will consist of multiple layers - a max layer for Pacman, and a min layer for each ghost. For instance, for a game with 3 ghosts, a search of depth 1 will consists of 4 levels in the search tree - one for Pacman, and then one for each of the ghosts.
Assignment 2, University of Toronto, CSC384 - Intro to AI,  
6

Score the leaves of your minimax tree with the supplied self.evaluationFunction, which defaults to scoreEvaluationFunction. You shouldn’t change this function now, but recognize that now we’re eval-uating *states* rather than actions, as we were for the reflex agent. Look-ahead agents evaluate future states whereas reflex agents evaluate actions from the current state.

You will have to implement a depth-bound specified in self.depth, so the leaves of your minimax tree could be either terminal or non-terminal nodes. Terminal nodes are nodes where either gameState.isWin() or gameState.isLose() is true. However, the leaves of your tree search might also be non-terminal nodes. You can run your MinimaxAgent on a game of Pacman by running the following command:

python pacman.py -p MinimaxAgent -l minimaxClassic -a depth=3

To test with the autograder, you may run:

python autograder.py -q q2

Hints and Observations

    • The correct implementation of minimax will lead to Pacman losing the game in some tests. This is not a problem: as it is correct behaviour, it will pass the tests.

    • Pacman is always agent 0, and the agents move in order of increasing agent index.

    • All states in minimax should be GameStates, either passed in to getAction or generated via GameS-tate.generateSuccessor. In this project, you will not be abstracting to simplified states.
    • On larger boards such as openClassic and mediumClassic (the default), you’ll find Pacman to be good at not dying, but quite bad at winning. He’ll often thrash around without making progress. Don’t worry if you see this behavior, Q4 and Q5 will clean up these issues.

Explorative questions – to submit (2 points):

    1. Note that Pacman will have suicidal tendencies when playing in situations where death is imminent. Why do you think this is the case? Briefly explain in one or two sentences. (2 points total)

Question 3: Alpha-Beta Pruning (10 points)

You will now implement a new agent that uses alpha-beta pruning to more efficiently explore the minimax tree. The function getAction in AlphaBetaAgent is found in multiAgents.py. Again, your algorithm must use the depth-bound specified in self.depth and evaluate its leaf nodes with self.evaluationFunction. Watch your agent play by running the following command:

python pacman.py -p AlphaBetaAgent -a depth=3 -l smallClassic

To test and debug your code, run

python autograder.py -q q3
Assignment 2, University of Toronto, CSC384 - Intro to AI,  
7

The correct implementation of alpha-beta pruning will lead to Pacman losing the game in some of the tests.

This is not a problem: as it is correct behaviour, it will pass the tests.

Explorative Questions – to submit (4 points total):

    1. You should notice a speed-up compared to your MinimaxAgent. Consider a game tree constructed for our Pacman game, where b is the branching factor and where depth is greater than d. Say a minimax agent (without alpha-beta pruning) has time to explore all game states up to and including those at level d. At level d, this agent will return estimated minimax values from the evaluation function.

        (a) In the best case scenario, to what depth would alpha-beta be able to search in the same amount of time? (1 point total)

        (b) In the worst case scenario, to what depth would alpha-beta be able to search in the same amount of time? How might this compare with the minimax agent without alpha-beta pruning? (2 points total)

    2. True or False: Consider a game tree where the root node is a max agent, and we perform a minimax search to terminals. Applying alpha-beta pruning to the same game tree may alter the minimax value of the root node. (1 point total)

Question 4: Expectimax (10 points)

Minimax and alpha-beta are great, but they both assume that you are playing against an adversary who makes optimal decisions. As anyone who has ever won tic-tac-toe can tell you, this is not always the case. In this question you will implement the ExpectimaxAgent, which is useful for modeling probabilistic be-havior of agents who may make suboptimal choices. The only difference in implementation of Expectimax search and Minimax search, is that at a min node, Expectimax search will return the average value over its children as opposed to the minimum value.

As with the search and constraint satisfaction problems covered in this class, the beauty of these algorithms is their general applicability. To expedite your own development, we’ve supplied some test cases based on generic trees. You can debug your implementation on small the game trees using the command:

python autograder.py -q q4

Debugging on these small and manageable test cases is recommended and will help you to find bugs quickly. Make sure when you compute your averages that you use floats. Integer division in Python truncates, so that 1/2 = 0, unlike the case with floats where 1.0/2.0 = 0.5.

To see how the ExpectimaxAgent behaves in Pacman, run:

python pacman.py -p ExpectimaxAgent -l minimaxClassic -a depth=3

The correct implementation of Expectimax will lead to Pacman losing some games in some of the tests.

This is not a problem: as it is correct behaviour, it will pass the tests.
Assignment 2, University of Toronto, CSC384 - Intro to AI,  
8

Explorative Questions – to submit (6 points):

    1. Consider a game tree where the root node is a max node, and the minimax value of the tree is vM. Consider a similar tree where the root node is a max node, but each min node is replaced with a chance node, where the expectimax value of the game tree is vE . For each of the following, decide whether the statement is True or False and briefly explain in one or two sentences your answer.

        (a) True or False: vM is always less than or equal to vE . Explain your answer. (2 points)

        (b) True or False: If we apply the optimal minimax policy to the game tree with chance nodes, we are guaranteed to result in a payoff of at least vM. Explain your answer. (2 points)

        (c) True or False: If we apply the optimal minimax policy to the game tree with chance nodes, we are guaranteed a payoff of at least vE . Explain your answer. (2 points)

Question 5: Evaluation Function (12 points)

Write a better evaluation function for Pacman in the provided function, betterEvaluationFunction. Your evaluation function should evaluate states (in contrast to the function employed by your reflex agent, which evaluated actions). You may use any tools at your disposal for evaluation. With depth 2 search, your evaluation function should clear the smallClassic layout with one random ghost more than half the time and still run at a reasonable rate. (To get full credit, Pacman should be averaging around 1000 points when he’s winning.)
python autograder.py -q q5

Grading: the autograder will run your agent on the smallClassic layout 10 times. We will assign points to your evaluation function in the following way:

    • If you win at least once without timing out the autograder, you receive 1 points. Any agent not satisfying these criteria will receive 0 points.

    • +1 for winning at least 5 times, +2 for winning all 10 times

    • +1 for an average score of at least 500, +2 for an average score of at least 1000 (including scores on lost games)

    • +1 if your games take on average less than 30 seconds on the autograder machine. The autograder is run on the teach.cs machines which have a fair amount of resources, but your personal computer could be far less performant (netbooks) or far more performant (gaming rigs). You can use your teach.cs login to run your program on the teach.cs machines.

    • The additional points for average score and computation time will only be awarded if you win at least 5 times.

Hints and Observations

    • As for your reflex agent evaluation function, you may want to use the reciprocal of important values (such as distance to food) rather than the values themselves.
Assignment 2, University of Toronto, CSC384 - Intro to AI,  
9

    • One way you might want to write your evaluation function is to use a linear combination of features. That is, compute values for features about the state that you think are important, and then combine those features by multiplying them by different values and adding the results together. You might decide what to multiply each feature by based on how important you think it is.

HAVE FUN and GOOD LUCK!