Homework 1 Characterizing an Agent's Environment

Introduction to AI
Homework 1 (100 pts)
General Information

Notes:
● For the written questions (Problem 1 and Problem 2 Part 5), put your answers in a
pdf (with your name in the pdf) and submit to the corresponding assignment on
Gradescope.
● For the coding questions (Problem 2 Parts 1­4), submit your code file
firstname_lastname_agents.py to the corresponding assignment on canvas, and also
append the code to the end of your pdf.
 For any solutions that require programming,
please comment and format your code to make it readable. You should code your solutions
in Python 2, and your code should be able to run on CAEN.
Problem 1: Characterizing an Agent's Environment [28 points]
In class we talked about specifying a task environment via the PEAS description. The PEAS
description contains the following specifications: Performance Measure, Environment,
Actuators, and Sensors.
In addition, environments can be described in terms of the following properties:
● Fully/Partially observable
● Deterministic/Stochastic
● Episodic/Sequential
● Static/Dynamic
● Discrete/Continuous
● Single­agent/Multi­agent
Describe the following agents using (A) the PEAS description and also (B) a description of
the task environment. Justify your answers, briefly (no more than 1 or 2 sentences per
term). Each question is worth 4 points (2 for PEAS and 2 for task environment).
1. Crossword Puzzle Solver
2. Roomba (https://en.wikipedia.org/wiki/Roomba)
3. Netflix Automatic Recommender System
4. Home Alarm
5. Boston Dynamics BigDog (https://www.youtube.com/watch?v=cNZPRsrwumQ)
6. Siri (https://www.apple.com/ios/siri/)
7. Paro Therapeutic Robot (http://www.parorobots.com/)
Problem 2: Implementing an Automatic Farming Agent (72 pts)
Consider an automated farming robot, FA­bot. It operates in a grid world, similar to the
world that the vacuum robot inhabits. The FA­bot can do four things:
1. It can move around the farm
2. It can sense properties of its environment
3. It can water crops
4. It can weed
The environment of the FA­bot is broken into grid squares. Each square may need to be
weeded or it may need to be watered. The locations are assigned using grid positions (even
if not all squares are available). The FA­bot must weed all squares that contain weeds, and
water all squares that have insufficient water. See Figure 1 for an illustration.
Figure 1: Board Configuration – a square with a cactus indicates that the square needs to be watered, a square
with a green plant indicates that it has to be weeded, and the square number is in the upper left corner (top left is
position [0,0]).
Environments are described by the text files env#.txt, where # is the environment number
(e.g., env1.txt, env2.txt,...). The first line of the file specifies the number of rows and the
number of columns of a grid. The second line contains the following configuration
information, in order:
1. Starting agent row
2. Starting agent column
3. Direction agent is initially facing (north, south, east, west)
4. Initial amount of water
5. How much water is used per watering
6. Initial amount of power
7. Power used for watering
8. Power used for weeding
9. Power used for moving
10. Power used for sensing
The remaining rows describe the environment in which the FA­Bot finds itself. Each
square contains either two Boolean values (e.g., TF) or two xx’s (e.g., xx). Squares marked
xx cannot be traversed by the agent. The other squares describe the need for watering (if
the first Boolean value is true) and the need for weeding (if the second is true).
The following is an example of an environment file that corresponds to Figure 1:
3 3
0 1 north 100 1 100 1 1 1 1
xx FT xx
TT TF FF
xx xx TT
This means that the agent starts in square [0,1] facing north, has 100 units of water and
power, and uses 1 unit of water with each watering action in addition to the 1 unit of power
that any action consumes. If, for example, the agent sensed the characteristics of square [0,
1], it would learn that the square does not need to be watered (F), but does need to be
weeded (T).
The goal configuration for this world is:
xx FF xx
FF FF FF
xx xx FF
The agent is successful only if there are no squares left to weed or water before the agent
runs out of power, and no squares are left to water before the agent runs out of water. The
agent should also try to conserve as much water and power as possible.
Note: you are provided with code for reading in these files.
Part 1: Implementing a Simple Reflex Agent (5 pts)
Problem setup
You have decided to use your AI knowledge to build an agent that will allow you to water
and weed the farmland. You will first build a simple reflex agent for this task by completing
the “choose_action” function under the “simple_reflex” class in agents.py. This agent cannot
maintain any internal state variables, so its behavior is defined entirely by the current
percept. The agent will receive a sequence of percepts, and take an action after each
percept, until either success or failure is achieved. The actions and corresponding percepts
are listed below:
Actions Description Possible Resulting
Percepts
­ Beginning the simulation start
sense_water Sense whether the current square
needs watering
needs_watering,
does_not_need_watering
sense_weed Sense whether the current square
needs weeding
needs_weeding,
does_not_need_weeding
water Water the current square watering_succeeded
weed Weed the current square weeding_succeeded
move
Move in the direction the agent is
facing. If there is a barrier, the agent
remains where it is, but rotates
clockwise.
move_succeeded,
hit_barrier
The agent will receive the “start” percept at the beginning of a simulation. The simulator
will halt when it receives a success or failure signal and print the results.
You should write this agent to do as well as possible, given its circumstances, constraints,
and goals. (So there is at least one right answer, which will produce optimal output for this
agent; don’t just return the same action every time and call it good).
Part 2: Implementing a Reflex Agent with State (5 pts)
Problem setup
You have decided that your FA­bot can do better by reasoning about its environment. To
do this it needs state.
You will implement a reflex agent with state, under the state_reflex class in agents.py. It
will plan its movements using breadth­first search (implemented for you in world.py).
What can the agent do?
This new agent can move in any desired direction. Its other actions are the same as before.
The actions are described in the table below:
Actions Description Possible Resulting
Percepts
­ Beginning the simulation start
sense_water Sense whether the current square
needs watering
needs_watering,
does_not_need_watering
sense_weed Sense whether the current square
needs weeding
needs_weeding,
does_not_need_weeding
water Water the current square watering_succeeded
weed Weed the current square weeding_succeeded
move_north The agent moves a space north, or does
nothing if a barrier is hit
move_succeeded,
hit_barrier
move_east The agent moves a space east, or does
nothing if a barrier is hit
move_succeeded,
hit_barrier
move_south The agent moves a space south, or does
nothing if a barrier is hit
move_succeeded,
hit_barrier
move_west The agent moves a space west, or does
nothing if a barrier is hit
move_succeeded,
hit_barrier
Unlike the agent of Part 1, this agent can see the layout of the environment in advance
(though it does not know what needs to be watered and weeded until it senses a square).
The agent is initialized by searching over the world with breadth­first search for a shortest
path of movement through the environment. It stores this movement plan, and how many
moves it has made, as state variables. This agent may not use other state variables
besides the ones we have defined for you (“plan” and “position”).
(Again, you should write this agent to do as well as possible).
Part 3: Implementing a Randomized Simple Reflex Agent (5 pts)
Problem setup
You want to see whether adding randomness to an agent can cause it to outperform the
simple reflex agent, without using state or performing any search in advance.
What can the agent do?
This randomized agent, implemented under class random_reflex in agents.py, has the same
set of actions as in Part 2, but does not maintain any internal state, and chooses among
move_north, move_east, move_south, and move_west uniformly at random when it decides
to move. You may use the numpy.random.randint function for randomness.
Part 4: Improving the Deterministic Simple Reflex Agent (8 pts)
Problem Setup
You get to have some fun with this one. Suppose that your agent cannot know any
information about the environment in advance. You want to improve the simple reflex
agent from Part 1 without adding any randomness to your code. To do this, you allow the
agent to maintain state variables, though it cannot do any search on the structure of the
world or otherwise see the world in advance, as it could in Part 2.
Implement the class better_reflex_agent in agents.py, such that the resulting deterministic
agent strictly outperforms the agent from Part 1 on at least one environment, and does
no worse than that agent on any other environment, including the held­out ones. This may
require a bit more thinking than the other parts, but you have more flexibility.
What can the agent do?
This agent can initialize and maintain any number of state variables, store any amount of
prior percepts/actions, use any built­in python data structures, and perform any number
of additional computations necessary, as long as they are deterministic. It may only use
the actions available to the agent in Part 1: sense_water, sense_weed, water, weed,
move.
Provided Code
In the file world.py we provide you with fully implemented classes GridSquare, and World,
as well as several classes and functions for breadth­first search. The class World reads in an
environment file during initialization and provides an interface for performing actions,
getting percepts, and tracking the agent’s water and power levels.
The file simulator.py contains the fully implemented functions simulate_agent() and
run_simulations(). Simulate_agent() runs a particular agent on a particular world and
prints the outcome. Run_simulations() is the function that we will use to test your code.
You should read the code files to make sure you understand how the provided code works.
Code You Must Write
For parts 1, 2, 3, and 4, you must fill in the designated sections of the file agents.py by
implementing the choose_action function for the four agent classes. The choose_action
functions of the first three agent classes are very similar to each other, and should not take
many lines of code.
How Simulations Work
After the world is initialized, agents are created and initialized. Each simulation begins by
feeding the agent the percept “start” via agent.choose_action(), and obtaining the agent’s
resulting action. The action is performed on the world and results in another percept,
which is again fed into the agent. This process continues until the simulator receives either
a “success” or “fail” percept, at which point it will halt and print the results. Your agent does
not need to handle the “success” or “fail” signals. The world maintains the agent’s position,
direction, and power and water levels, so you don’t have to.
How we will run your code
We will test your code on some small held­out environments, mainly to make sure that it
runs and that it does what it’s supposed to. We will import your (renamed) version of
agents.py into simulator.py and run
python simulator.py <firstname_lastname_agent <env_name.txt <agenttype
from the command line, for each environment and agent type, where
“firstname_lastname_agent” is the name of your agent file without the file extension,
env_name.txt is an environment file, and “agenttype” is one of simple_reflex, state_reflex,
and random_reflex. For example,
python simulator.py john_doe_agent env1.txt simple_reflex
You can test your code with these commands. Add ‘v’ to the end of the line for verbose
output.
Your code’s performance here is worth 6 points.
Part 5: Written Questions for Problem 2
For Problem 2, include answers to the following questions in your pdf. Questions are worth
4 pts each, unless otherwise specified.
1. Write a PEAS description for FA­bot’s environment in Part 1.
2. What is the definition of rational behavior in FA­bot’s environment?
3. Can the simple reflex agent from Part 1 be perfectly rational for this environment?
What about the agent from Part 2? Why or why not?
4. For each of parts 1, 2, 3, and 4, fill in a table like the following that contains the
outcomes of each agent running on each provided environment. (For the
randomized agent, there is, of course, no single correct answer. Provide averages
over 5 runs, with percent of times succeeded in the “succeeded” column).
Env. # Succeeded? # Moves Water Power # Left to water # Left to weed
5. Which agent performed the best overall? Why? Does your randomized simple reflex
agent from Part 3 outperform your deterministic agent from Part 1 in any
environments, and why? List the ones that it does, if any.
6. (7 pts) In which environment(s) did your agent from Part 4 outperform the agent
from Part 1? Describe how you structured the better reflex agent, and why it
performs better in some cases and no worse in others. (Or, if you did not succeed in
making such an agent, describe what you were trying to do).
Resource constraints lead to trade­offs between different strategies that agents could
implement. Consider the following simplified scenario: an agent must move along a single
row of grid squares, which either do or don’t contain weeds with about 50% probability
(e.g., the grid looks like F T T F T …). The agent can move forward, sense for weeds, and
weed (like our agents above, the weed action can be performed whether or not there is
actually a weed). The agent wants to weed all squares while using as little power as
possible.
7. Describe this agent’s optimal strategy assuming it costs 1 unit of power to sense for
weeds and 5 units of power to weed.
8. How does this agent’s strategy change if it costs 5 units of power to sense for weeds
and 1 unit of power to weed?
The internal logic of our agents above was greatly simplified by environmental
determinism. We knew that if the agent attempted a “weed” action, for example, the action
would succeed with certainty. But, what happens when certainty is no longer guaranteed?
Assume that we are working with the state­based reflex FA­bot that you implemented from
Part 2.
9. Now our FA­bot is on the fritz. We have been meaning to replace the agent, but, we
haven’t. As a result, the “weed” action only succeeds in removing weeds about 75%
of the time! How would the performance of the state­based agent from Part 2 be
affected by this indeterminism?
10. Describe how you would you modify that agent to handle this sort of indeterminism.
Rules for Programming and Submission
● After you fill out agents.py, rename it to firstname_lastname_agents.py and submit
it in canvas under assignment 1. (Of course, “firstname” and “lastname” should be
replaced with your actual first and last name).
● Submit a neat pdf of answers to the questions (with your code pasted at the end) to
Gradescope. Be sure to identify the correct pages with the correct problems. Put
your name in the pdf.
● Include a comment at the top of your source file with your name
● Your program must run from a command line on CAEN computers.
● Use good style and layout. Comment your code well.
Answer Format and Grading
● Answers to questions should be complete but concise ­ don’t be more wordy than
necessary. Most questions can be answered in a couple of sentences at the most.
● Partial or (possibly) full credit will be given to answers that are well­explained, even
if they are different from our answers. (Or code that is well­written, even if it has
errors).
● Points are broken down as follows:
○ 28 pts for Problem 1, or 4 pts per subproblem
○ 5 pts for the agent code of Problem 2 Part 1
○ 5 pts for the agent code of Problem 2 Part 2
○ 5 pts for the agent code of Problem 2 Part 3
○ 8 pts for the agent code of Problem 2 Part 4
○ 6 pts for all your agent code running and producing good output on our test
environments
○ 43 pts for Problem 2, Part 5, 4 pts per subproblem, with 7pts for Q6.