共计 6157 个字符，预计需要花费 16 分钟才能阅读完成。
COMP3211 2021-22 Spring Semester Project on Multi-Agent Path Finding
(For Individuals)
Date assigned: Thursday, Mar 31.
Due time: 23:59 on Tuesday, May 10.
How to submit it: Submit your work as a zip file – see below for details.
Penalties on late submission: 20% off each day (anytime after the due time is considered
late by one day)
This document is for those who choose to do the project individually. The tasks are exactly
the same as the ones for the team project except that you only need to consider two agents.
1 Project description
Imagine in a warehouse, a fleet of robots are designed to pick up packages from certain
locations and deliver them to certain ports. They should deliver the packages as efficiently
as possible, and at the same time, must not bump into each other or obstacles (no collision).
In this project, you are going to design environment specific controllers for these robots.
There are three environments called small, medium and large – see the appendix. Each
environment comes with two fixed goal positions, one for each robot.
For each environment, you need to come up with 2 policies (pi1, pi2), one for each robot.
A policy for a robot is a function from states to actions, with the following definitions of
states and actions:
A state is a tuple (l1, l2), where li is the current location of robot i.
The actions are: up, down, right, left, nil, with their usual meaning.
We will test your policies on some random initial states, and measure the quality of them
by the sum of the numbers of actions the robots take to arrive at their respective goal
locations.
2 Project details
This project comes with a simulator and a code skeleton.
2.1 File Description
- game.py: provides the class Env which takes as input a map and a set of goals, and
the class Game which takes as input a set of initial positions, a group of agents and an
1
instantiated environment. - base.py: provides several helper functions as well as the BaseAgent class which re-
stricts the inputs of an agent to only include a name and an instantiated environment. - run.py: parses the given commandline. You can also see how the agent will be tested
in the main function. - animator.py: visualizes the whole process, you do not have look into this file.
- map/: provides three maps named small.map, medium.map, large.map. You agent
will only be tested on these three given maps (i.e., no hidden map will be used for
evaluation). Each map file is also associated with a set of goals and a constant
MAX_NUM_STEP representing the maximum number of steps allowed in this map. - agent.py: a simple demo agent.
2.2 Environment Set-up - Check if anaconda has been installed on your PC,
conda -V - Create a virtual environment for this project with python 3.8 and activate the envi-
ronment,
conda create -n mapf python=3.8
conda activate mapf - Install dependency packages,
pip install -r requirements.txt
2.3 Run the Code - Activate the environment,
conda activate mapf - You are given a simple agent in agent.py, to run it,
python run.py –agents p1 p2 –map empty –goals 5_5 1_5 –vis
This command launched two agents named‘p1’and‘p2’on the map named‘empty’
and goals are specified as (5,5) and (1,5) for each, respectively. The –vis option
turns on the visualizer. Note that the visualizer for large maps is quite slow, you may
want to turn off the –vis option for large maps.
2 - Specify the initial positions for each agent as input,
Specify an initial position for agent p1: 1 1
Specify an initial position for agent p2: 5 1
If the code runs successfully, then you will see an animation window like this as well
as the whole path finding history printed in the terminal.
Note that you only need to specify the map and goals once but then different initial
positions are allowed to input as many times as you like, which supports what we
decribed as policies above. If you want to terminate the program, just input n as an
invalid coordinate.
Specify an initial position for agent p1: n - For further detailed usage,
python run.py -h
2.4 Tasks
Your main task is to implement MyAgent class in agent.py. In particular, you need to
implement a MyAgent.get_action() function. This function is basically your controller:
it will be called to decide which action to do in each state. In order for the robot to run in
realtime, a hard constraint is that every call to this function will need to return in - second. Otherwise the nil action is selected for the agent. This function will have access
to information about which environment the robots are in so that your controller can be
environment specific.
You are allowed to 1) implement any other helper functions, which can be global functions
or member functions; 2) import any other packages. This will enable you to design your
controller in any way you want. For example, you can build a database of state-action pairs.
3
Then your MyAgent.get_action() can simply perform query to this database. However,
this is unlikely to be effective for the large map as the number of states is very large.
Another approach is to train a machine learning model for each map, and then use this
model in your controller.
2.5 Evaluation
For the evaluation of each map, we will instantiate a team of your agents once and then test
your agents with 10/20/30 (for maps of different sizes) randomly generated sets of initial
positions. For each set of initial positions, you will have a time limit of MAX_NUM_STEP.
Let total_num_steps be the sum of the numbers of actions the robots take to arrive their
respective goal locations, your score in that round will be?????
100? 100 · total_num_stepsMAX_NUM_STEP , all agents accomplish their jobs within the time limit - , any collision happens
- , otherwise
Your score for that particular map will be the average of scores over all rounds. Finally,
your total score will be
15%·avg_score(small_map)+25%·avg_score(medium_map)+40%·avg_score(large_map)
As mentioned above, you only need to consider two-agent cases. That is, your submitted
agent will be tested using the following command,
python run.py –agents p1 p2 –map {map} –goals {g1} {g2} –eval - Submission
You must submit a report of your project (report.pdf, one column, at most 3 pages). It
must describe the methods that you used to design your controller.
Before submission, please re-write the dependency file if you have used additional packages
(remember to switch to the project directory first),
pip install pipreqs
pipreqs –force ./
4
Then arrange your files as follows in a folder named as your studentID and submit the whole
zip file (e.g., 123456.zip).
{studentID}/
|- report.pdf
|- agent.py
|- requirements.txt
|- extra/
|- other_useful_file1
|- …
You can put any of your trained model files or helper functions under the extra/ folder.
20% of your grade will be based on your report and 80% from the total score that your
controllers get in evaluation as described above. - Enquiries
Enquiries before deadline: Fengming ZHU (fzhuae@connect.ust.hk)
Grading:
– Zhili CHEN (zchenei@connect.ust.hk),
– Chenglin WANG (chenglin.wang@connect.ust.hk) - Appendix – maps
A small map (Figure 1, size: 8× 8):
A medium-sized map (Figure 2, size: 18× 18):
A large map (Figure 3, part of Berlin, size: 256× 256):
5
Figure 1: goal1(orange) = (5, 5), goal2(blue) = (3, 3)
Figure 2: goal1(orange) = (9, 7), goal2(blue) = (5, 9)
6
Figure 3: goal1(orange) = (150, 125), goal2(blue) = (100, 175)