Difference between revisions of "MC3-Project-2"

From Quantitative Analysis Software Courses
Jump to navigation Jump to search
 
(123 intermediate revisions by 2 users not shown)
Line 1: Line 1:
 
==Updates / FAQs==
 
==Updates / FAQs==
  
* Q: In a previous project there was a constraint of holding a single position until exit. Does that apply to this project?  Yes, hold one position til exit.
+
* '''2017-03-29''' Changed rubric to evaluate the '''median''' of all runs rather than the "best" of all runs.
 +
* '''2016-11-8''' Changed rubric to evaluate the '''median''' of all runs rather than the "best" of all runs.
  
* Q: Is that 5 calendar days, or 5 trading days (i.e., days when SPY was traded)? A: Always use trading days.
+
==Overview==
 +
 
 +
In this project you will implement the Q-Learning and Dyna-Q solutions to the reinforcement learning problem.  You will apply them to a navigation problem in this project.  In a later project you will apply them to trading.  The reason for working with the navigation problem first is that, as you will see, navigation is an easy problem to work with and understand.  Note that your Q-Learning code really shouldn't care which problem it is solving.  The difference is that you need to wrap the learner in different code that frames the problem for the learner as necessary.
 +
 
 +
For the navigation problem we have created testqlearner.py that automates testing of your Q-Learner in the navigation problem.  
 +
 
 +
Overall, your tasks for this project include:
 +
 
 +
* Code a Q-Learner
 +
* Code the Dyna-Q feature of Q-Learning
 +
* Test/debug the Q-Learner in navigation problems
  
* Q: Are there constraints for Python modules allowed for this project? Can we experiment with modules for optimization or technical analysis and cite or are we expected to write everything from scratch for this project as well?  A: You can use scikit modules as long as you cite them..  You've already written the learners you need though.
+
For this assignment we will test only your code (there is no report component).
  
* Q: Can we change our policy to work better for IBM vs the sine data? A: No, you must use the same indicators, policy, etc. for both.  I suggest you optimize first for IBM, then go back to the sine data because almost anything should work with the sine data.
+
==Template and Data==
  
==Overview==
+
* Update your local mc3p2_qlearning_robot directory using github.
 +
* Implement the <tt>QLearner</tt> class in <tt>mc3p2_qlearning_robot/QLearner.py</tt>.
 +
* To debug your Q-learner, run <tt>'''python testqlearner.py'''</tt> from the <tt>mc3p2_qlearning_robot/</tt> directory. The grading script for this project is <tt>grade_robot_qlearning.py</tt>.
 +
* Note that example navigation problems are provided in the <tt>mc3p2_qlearning_robot/testworlds</tt> directory.
  
In this project you will transform your regression learner into a stock trading strategy.  You should train a learner to predict the change in price of a stock over the next five trading days (one week). You will use data from Dec 31 2007 to 2009 to train your prediction model, then you will test it from Dec 31 2009 to 2011.
+
==Part 1: Implement Q-Learner (95%)==
  
Now, just predicting the change in price isn't enough, you need to also code a policy that uses the forecaster you built to buy or sell sharesYour policy should buy when it thinks the price will go up, and short when it thinks the price will go downYou can then feed those buy and sell orders into your market simulator to backtest the strategy.
+
Your QLearner class should be implemented in the file <tt>QLearner.py</tt>It should implement EXACTLY the API defined belowDO NOT import any modules besides those allowed below.  Your class should implement the following methods:
  
Finding features, a learner, and a policy that all work together to provide a reliably winning strategy is HARD! It is possible, and people have done it, but we can't reasonably expect you to be successful at it in this short class.  Accordingly, we want you to work with some easy data first, namely we will provide you with sinusoidal historical price dataOnce you've got something that works with that, you can try your learner on real stock data.
+
<b>The constructor QLearner()</b> should reserve space for keeping track of Q[s, a] for the number of states and actions. It should initialize Q[] with all zerosDetails on the input arguments to the constructor:
  
==Detailed steps==
+
* <tt>num_states</tt> integer, the number of states to consider
 +
* <tt>num_actions</tt>  integer, the number of actions available.
 +
* <tt>alpha</tt> float, the learning rate used in the update rule. Should range between 0.0 and 1.0 with 0.2 as a typical value.
 +
* <tt>gamma</tt> float, the discount rate used in the update rule.  Should range between 0.0 and 1.0 with 0.9 as a typical value.
 +
* <tt>rar</tt> float, random action rate: the probability of selecting a random action at each step. Should range between 0.0 (no random actions) to 1.0 (always random action) with 0.5 as a typical value.
 +
* <tt>radr</tt> float, random action decay rate, after each update, rar = rar * radr. Ranges between 0.0 (immediate decay to 0) and 1.0 (no decay).  Typically 0.99.
 +
* <tt>dyna</tt> integer, conduct this number of dyna updates for each regular update.  When Dyna is used, 200 is a typical value.
 +
* <tt>verbose</tt> boolean, if True, your class is allowed to print debugging statements, if False, all printing is prohibited.
  
Overall, you should follow these steps:
+
<b>query(s_prime, r)</b> is the core method of the Q-Learner.  It should keep track of the last state s and the last action a, then use the new information s_prime and r to update the Q table.  The learning instance, or experience tuple is <s, a, s_prime, r>.  query() should return an integer, which is the next action to take.  Note that it should choose a random action with probability rar, and that it should update rar according to the decay rate radr at each step.  Details on the arguments:
  
* Train a regression learner (KNN or LinReg, or other of your choice with or without bagging) on data from Dec 31 2007 to Dec 31 2009.  This is your in sample training data.
+
* <tt>s_prime</tt> integer, the the new state.
** For your X values: Identify and implement at least 3 technical features that you believe may be predictive of future return. You should implement them so they output values typically ranging from -1.0 to 1.0.  This will help avoid the situation where one feature overwhelms the results. See a few formulae below.
+
* <tt>r</tt> float, a real valued immediate reward.
** For your Y values: Use future 5 day return (not future price).  You're trying to predict a relative change that you can use to invest with.
 
* Create a plot that illustrates your training Y values in one color, current price in another color and your model's PREDICTED Y in a third color. To help with the visualization, you should adjust your training Y and predicted Y so that they are at the same scale as the current price. With this chart we should be able to see how well your learner performs and that your Y values are shifted back 5 days.  You may find it convenient to zoom in on a particular time period so this is evident.
 
* Create a trading policy based on what your learner predicts for future return.  As an example you might choose to buy when the forecaster predicts the price will go up more than 1%, then hold for 5 days.
 
* Create a plot that illustrates entry and exits as vertical lines on a price chart for the in sample period Dec 31 2007 to Dec 31 2009. Show long entries as green lines, short entries as red lines and exits as black lines. You may find it convenient to zoom in on a particular time period so this is evident. 
 
* Now use your code to generate orders and run those orders through your market simulator.  Create a chart of this backtest.  It should do VERY well for the in sample period Dec 31 2007 to Dec 31 2009.
 
* Freeze your model based on the Dec 31 2007 to Dec 31 2009 training data.  Now test it out of sample over the period Dec 31 2009 to Dec 31 2011.  Create a plot that illustrates entry & exits, generate trades, run through your simulator, chart the backtest.
 
  
Perform the above steps first using the data ML4T-220.csv.  Once you've validated success (it should work well), repeat using IBM data over the same datesRemember Dec 31 2007 to Dec 31 2009 is training, Dec 31 2009 to Dec 31 2011 is testing. You should have one set of charts for each symbol.
+
<b>querysetstate(s)</b> A special version of the query method that sets the state to s, and returns an integer action according to the same rules as query() (including choosing a random action sometimes), but it does not execute an update to the Q-tableIt also does not update rar. There are two main uses for this method: 1) To set the initial state, and 2) when using a learned policy, but not updating it.
  
==Summary of Plots To Create==
+
Here's an example of the API in use:
  
# Training Y/Price/Predicted Y: Create a plot that illustrates your training Y values in one color, current price in another color and your model's PREDICTED Y in a third color. To help with the visualization, you should adjust your training Y and predicted Y so that it is at the same scale as the current price.
+
<PRE>
# Sine data in-sample Entries/Exits: Create a plot that illustrates entry and exits as vertical lines on a price chart for the in sample period. Show long entries as green lines, short entries as red lines and exits as black lines. You may find it convenient to zoom in on a particular time period so this is evident.
+
import QLearner as ql
# Sine data in-sample backtest
 
# Sine data out-of-sample Entries/Exits: Freeze your model based on the in-sample data. Now test it for the the out-of-sample period. Plot the entry & exits, generate trades,
 
# Sine data out-of-sample backtest.
 
# IBM data in-sample Entries/Exits: Create a plot that illustrates entry and exits as vertical lines on a price chart for the in sample period 2008-2009. Show long entries as green lines, short entries as red lines and exits as black lines. You may find it convenient to zoom in on a particular time period so this is evident.
 
# IBM data in-sample backtest
 
# IBM data out-of-sample Entries/Exits
 
# IBM data out-of-sample backtest
 
  
==Template and Data==
+
learner = ql.QLearner(num_states = 100, \
 +
    num_actions = 4, \
 +
    alpha = 0.2, \
 +
    gamma = 0.9, \
 +
    rar = 0.98, \
 +
    radr = 0.999, \
 +
    dyna = 0, \
 +
    verbose = False)
 +
 
 +
s = 99 # our initial state
  
You should create a directory for your code in ml4t/mc3-p2.  You will have access to the data in the ML4T/Data directory but you should use ONLY the code in util.py to read it.  In particular files named ML4T-220.csv, and IBM.csv.
+
a = learner.querysetstate(s) # action for state s
  
==Choosing Technical Features -- Your X Values==
+
s_prime = 5 # the new state we end up in after taking action a in state s
  
You should have already successfully coded the Bollinger Band feature.  Here's a suggestion of how to normalize that feature so that it will typically provide values between -1.0 and 1.0:
+
r = 0 # reward for taking action a in state s
  
<PRE>
+
next_action = learner.query(s_prime, r)
bb_value[t] = (price[t] - SMA[t])/(2 * stdev[t])
 
 
</PRE>
 
</PRE>
  
Two other good features worth considering are momentum and volatility.
+
==Part 2: Navigation Problem Test Cases==
 +
 
 +
We will test your Q-Learner with a navigation problem as follows.  Note that your Q-Learner does not need to be coded specially for this task.  In fact the code doesn't need to know anything about it.  The code necessary to test your learner with this navigation task is implemented in testqlearner.py for you.  The navigation task takes place in a 10 x 10 grid world.  The particular environment is expressed in a CSV file of integers, where the value in each position is interpreted as follows:
 +
 
 +
* 0: blank space.
 +
* 1: an obstacle.
 +
* 2: the starting location for the robot.
 +
* 3: the goal location.
 +
* 5: quicksand.
 +
 
 +
An example navigation problem (world01.csv) is shown below.  Following python conventions, [0,0] is upper left, or northwest corner, [9,9] lower right or southeast corner.  Rows are north/south, columns are east/west.
  
 
<PRE>
 
<PRE>
momentum[t] = (price[t]/price[t-N]) - 1
+
3,0,0,0,0,0,0,0,0,0
 +
0,0,0,0,0,0,0,0,0,0
 +
0,0,0,0,0,0,0,0,0,0
 +
0,0,1,1,1,1,1,0,0,0
 +
0,5,1,0,0,0,1,0,0,0
 +
0,5,1,0,0,0,1,0,0,0
 +
0,0,1,0,0,0,1,0,0,0
 +
0,0,0,0,0,0,0,0,0,0
 +
0,0,0,0,0,0,0,0,0,0
 +
0,0,0,0,2,0,0,0,0,0
 
</PRE>
 
</PRE>
  
Volatility is just the stdev of daily returns.
+
In this example the robot starts at the bottom center, and must navigate to the top left.  Note that a wall of obstacles blocks its path, and there is some quicksand along the left side.  The objective is for the robot to learn how to navigate from the starting location to the goal with the highest total reward.  We define the reward for each step as:
 +
* -1 if the robot moves to an empty or blank space, or attempts to move into a wall
 +
* -100 if the robot moves to a quicksand space
 +
* 1 if the robot moves to the goal space
 +
 
 +
Overall, we will assess the performance of a policy as the median reward it incurs to travel from the start to the goal (higher reward is better).  We assess a learner in terms of the reward it converges to over a given number of training epochs (trips from start to goal).  '''Important note:''' the problem includes random actions.  So, for example, if your learner responds with a "move north" action, there is some probability that the robot will actually move in a different direction.  For this reason, the "wise" learner develops policies that keep the robot well away from quicksand.  We map this problem to a reinforcement learning problem as follows:
 +
 
 +
* State: The state is the location of the robot, it is computed (discretized) as: column location * 10 + row location.
 +
* Actions: There are 4 possible actions, 0: move north, 1: move east, 2: move south, 3: move west.
 +
* R: The reward is as described above.
 +
* T: The transition matrix can be inferred from the CSV map and the actions.
  
==Choosing Y==
+
Note that R and T are not known by or available to the learner.  The code in <tt>testqlearner.py</tt> will test your code as follows (pseudo code):
  
Your code should predict 5 day change in price.  You need to build a new Y that reflects the 5 day change and aligns with the current date. Here's pseudo code for the calculation of Y
+
<pre>
 +
Instantiate the learner with the constructor QLearner()
 +
s = initial_location
 +
a = querysetstate(s)
 +
s_prime = new location according to action a
 +
r = -1.0
 +
while not converged:
 +
    a = query(s_prime, r)
 +
    s_prime = new location according to action a
 +
    if s_prime == goal:
 +
        r = +1
 +
        s_prime = start location
 +
    else if s_prime == quicksand:
 +
        r = -100
 +
    else:
 +
        r = -1
 +
</pre>
  
Y[t] = (price[t+5]/price[t]) - 1.0
+
A few things to note about this code: The learner always receives a reward of -1.0 (or -100.0) until it reaches the goal, when it receives a reward of +1.0. As soon as the robot reaches the goal, it is immediately returned to the starting location.
  
If you select Y in this manner and use it for training, your learner will predict 5 day returns.
+
Here are example solutions.  Note that these examples were created before we added "quicksand" to the project.  We will be updating the examples to reflect this change.  In the mean time, you may find these useful:
 +
 
 +
[[mc3_p2_examples]]
 +
 
 +
[[mc3_p2_dyna_examples]]
 +
 
 +
==Part 3: Implement Dyna (5%)==
 +
 
 +
Add additional components to your QLearner class so that multiple "hallucinated" experience tuples are used to update the Q-table for each "real" experience.  The addition of this component should speed convergence in terms of the number of calls to query(), <b>not necessarily running time</b>.
 +
 
 +
Note that it is not important that you implement Dyna exactly as described in the lecture.  The key requirement is that your code should somehow hallucinate additional experiences.  The precise method you use for discovering those experiences is flexible.  We will test your code on several test worlds with 50 epochs and with dyna = 200.  Our expectation is that with Dyna, the solution should be much better after 50 epochs than without.
 +
 
 +
==Part 4: Implement author() Method (0%)==
 +
 
 +
You  should implement a method called <tt>author()</tt> that returns your Georgia Tech user ID as a string. This is the ID you use to log into t-square.  It is not your 9 digit student number.  Here is an example of how you might implement author() within a learner object:
 +
 
 +
<pre>
 +
class QLearner(object):
 +
    def author(self):
 +
        return 'tb34' # replace tb34 with your Georgia Tech username.
 +
</pre>
 +
 
 +
And here's an example of how it could be called from a testing program:
 +
 
 +
<pre>
 +
    # create a learner and train it
 +
    learner = ql.QLearner() # create a QLearner
 +
    print learner.author()
 +
</pre>
 +
 
 +
Check the template code for examples. We are adding those to the repo now, but it might not be there if you check right away.  Implementing this method correctly does not provide any points, but there will be a penalty for not implementing it.
  
 
==Contents of Report==
 
==Contents of Report==
  
* Your report should be no more than 2500 words.  Your report should contain no more than 12 chartsPenalties will apply if you violate these constraints.
+
There is no report component of this assignmentHowever, if you would like to impress us with your Machine Learning prowess, you are invited to submit a succinct report.
* Include the charts listed in the overview section above.
 
* Describe each of the indicators you have selected in enough detail that someone else could reproduce them in code.
 
* Describe your trading policy clearly.
 
* Discussion of results. Did it work well?  Why?  What would you do differently?
 
  
 
==Hints & resources==
 
==Hints & resources==
 +
 +
This paper by Kaelbling, Littman and Moore, is a good resource for RL in general: http://www.jair.org/media/301/live-301-1562-jair.pdf  See Section 4.2 for details on Q-Learning.
 +
 +
There is also a chapter in the Mitchell book on Q-Learning.
 +
 +
For implementing Dyna, you may find the following resources useful:
 +
 +
* https://webdocs.cs.ualberta.ca/~sutton/book/ebook/node96.html
 +
* http://www-anw.cs.umass.edu/~barto/courses/cs687/Chapter%209.pdf
  
 
==What to turn in==
 
==What to turn in==
  
Turn your project in via t-square.
+
Turn your project in via t-square.   All of your code must be contained within QLearner.py .
  
* Your report as <tt>report.pdf</tt>
+
* Your QLearner as <tt>QLearner.py</tt>
* All of your code, as necessary to run as <tt>.py</tt> files.
+
* Do not submit any other files.
* Document how to run your code in <tt>readme.txt</tt>.
 
  
==Extra credit up to 3%==
+
==Rubric==
  
Choose one or more of the following:
+
Only your QLearner class will be tested. 
  
* Compare the performance of KNN and LinReg in this taskThe instructor anticipates that LinReg might work wellIf that turns out to be the case, how can that be?  This is a non-linear task.
+
* For basic Q-Learning (dyna = 0) we will test your learner against 10 test worlds with 500 epochs in each world.  One "epoch" means your robot reaches the goal one time, or after 100000 steps, whichever comes first.  Your QLearner retains its state (Q-table), and then we allow it to navigate to the goal again, over and over, 500 timesEach test (500 epochs) should complete in less than 2 seconds<b>NOTE</b>: an epoch where the robot fails to reach the goal will likely take <b>much</b> longer (in running time), than one that does reach the goal, and is a common reason for failing to complete test cases within the time limit.
* Extend your code to create a "rolling" model that updates each day rolling forward.
+
* Benchmark: As a benchmark to compare your solution to, we will run our reference solution in the same world, with 500 epochs.  We will take the median reward of our reference across all of those 500 epochs.
* Extend your code to simultaneously forecast all the members of the S&P 500. Generate trades accordingly, and backtest the result.
+
* Your score: For each world we will take the median cost your solution finds across all 500 epochs.
 +
* For a test to be successful, your learner should find a total reward >= 1.5 x the benchmark. Note that since reward for a single epoch is negative, your solution can be up to 50% worse than the reference solution and still pass.
 +
* There are 10 test cases, each test case is worth 9.5 points.
 +
* Here is how we will initialize your QLearner for these test cases:
  
Submit to the extra credit assignment on t-square. One single PDF file only, max 5 pages.
+
<Pre>
 +
    learner = ql.QLearner(num_states=100,\
 +
        num_actions = 4, \
 +
        alpha = 0.2, \
 +
        gamma = 0.9, \
 +
        rar = 0.98, \
 +
        radr = 0.999, \
 +
        dyna = 0, \
 +
        verbose=False) #initialize the learner
 +
</PRE>
  
==Rubric==
+
* For Dyna-Q, we will set dyna = 200.  We will test your learner against <tt>world01.csv</tt> and <tt>world02.csv</tt> with 50 epochs.  Scoring is similar to the non-dyna case: Each test should complete in less than 10 seconds. For the test to be successful, your learner should find solution with total reward to the goal >= 1.5 x the  median reward our reference solution across all 50 epochs.  Note that since reward for a single epoch is negative, your solution can be up to 50% worse than the reference solution and still pass.  We will check this by taking the median of all 50 runs. Each test case is worth 2.5 points.  We will initialize your learner with the following parameter values for these test cases:
 +
 
 +
<Pre>
 +
    learner = ql.QLearner(num_states=100,\
 +
        num_actions = 4, \
 +
        alpha = 0.2, \
 +
        gamma = 0.9, \
 +
        rar = 0.5, \
 +
        radr = 0.99, \
 +
        dyna = 200, \
 +
        verbose=False) #initialize the learner
 +
</PRE>
  
* Are all 9 plots present and correct? -5 points for each missing plot.
+
* Is the author() method correctly implemented (-20% if not)
** Note: Correct in the sense that they properly display the information requested.  The result may not be the desired one.
 
* Are numerical results correct? (ML4T-220 in sample & out of sample, IBM in sample & out of sample) -10 points for each incorrect result.
 
* Indicators used: Are descriptions of factors used sufficiently clear that others could reproduce them? Up to -10 points for lack of clarity.
 
* Trading strategy: Is description sufficiently clear that others could reproduce it? Up to -10 points for lack of clarity.
 
* Is discussion of results concise, complete, correct? Up to -5 points for each of concise, complete, correct.
 
  
 
==Required, Allowed & Prohibited==
 
==Required, Allowed & Prohibited==
 
[for 2016]
 
  
 
Required:
 
Required:
 
* Your project must be coded in Python 2.7.x.
 
* Your project must be coded in Python 2.7.x.
 +
* Your code must run on one of the university-provided computers (e.g. buffet02.cc.gatech.edu).
  
 
Allowed:
 
Allowed:
Line 127: Line 239:
 
* Any code you did not write yourself (except for the 5 line rule in the "allowed" section).
 
* Any code you did not write yourself (except for the 5 line rule in the "allowed" section).
 
* Any Classes (other than Random) that create their own instance variables for later use (e.g., learners like kdtree).
 
* Any Classes (other than Random) that create their own instance variables for later use (e.g., learners like kdtree).
* Holy hand grenades.
+
* Print statements outside "verbose" checks (they significantly slow down auto grading).
 +
* Any method for reading data besides util.py
 +
 
 +
==Legacy==
 +
 
 +
*[[MC3-Project-2-Legacy-trader]]
 +
*[[MC3-Project-2-Legacy]]
 +
*[[MC3-Project-4]]

Latest revision as of 14:37, 3 July 2017

Updates / FAQs

  • 2017-03-29 Changed rubric to evaluate the median of all runs rather than the "best" of all runs.
  • 2016-11-8 Changed rubric to evaluate the median of all runs rather than the "best" of all runs.

Overview

In this project you will implement the Q-Learning and Dyna-Q solutions to the reinforcement learning problem. You will apply them to a navigation problem in this project. In a later project you will apply them to trading. The reason for working with the navigation problem first is that, as you will see, navigation is an easy problem to work with and understand. Note that your Q-Learning code really shouldn't care which problem it is solving. The difference is that you need to wrap the learner in different code that frames the problem for the learner as necessary.

For the navigation problem we have created testqlearner.py that automates testing of your Q-Learner in the navigation problem.

Overall, your tasks for this project include:

  • Code a Q-Learner
  • Code the Dyna-Q feature of Q-Learning
  • Test/debug the Q-Learner in navigation problems

For this assignment we will test only your code (there is no report component).

Template and Data

  • Update your local mc3p2_qlearning_robot directory using github.
  • Implement the QLearner class in mc3p2_qlearning_robot/QLearner.py.
  • To debug your Q-learner, run python testqlearner.py from the mc3p2_qlearning_robot/ directory. The grading script for this project is grade_robot_qlearning.py.
  • Note that example navigation problems are provided in the mc3p2_qlearning_robot/testworlds directory.

Part 1: Implement Q-Learner (95%)

Your QLearner class should be implemented in the file QLearner.py. It should implement EXACTLY the API defined below. DO NOT import any modules besides those allowed below. Your class should implement the following methods:

The constructor QLearner() should reserve space for keeping track of Q[s, a] for the number of states and actions. It should initialize Q[] with all zeros. Details on the input arguments to the constructor:

  • num_states integer, the number of states to consider
  • num_actions integer, the number of actions available.
  • alpha float, the learning rate used in the update rule. Should range between 0.0 and 1.0 with 0.2 as a typical value.
  • gamma float, the discount rate used in the update rule. Should range between 0.0 and 1.0 with 0.9 as a typical value.
  • rar float, random action rate: the probability of selecting a random action at each step. Should range between 0.0 (no random actions) to 1.0 (always random action) with 0.5 as a typical value.
  • radr float, random action decay rate, after each update, rar = rar * radr. Ranges between 0.0 (immediate decay to 0) and 1.0 (no decay). Typically 0.99.
  • dyna integer, conduct this number of dyna updates for each regular update. When Dyna is used, 200 is a typical value.
  • verbose boolean, if True, your class is allowed to print debugging statements, if False, all printing is prohibited.

query(s_prime, r) is the core method of the Q-Learner. It should keep track of the last state s and the last action a, then use the new information s_prime and r to update the Q table. The learning instance, or experience tuple is <s, a, s_prime, r>. query() should return an integer, which is the next action to take. Note that it should choose a random action with probability rar, and that it should update rar according to the decay rate radr at each step. Details on the arguments:

  • s_prime integer, the the new state.
  • r float, a real valued immediate reward.

querysetstate(s) A special version of the query method that sets the state to s, and returns an integer action according to the same rules as query() (including choosing a random action sometimes), but it does not execute an update to the Q-table. It also does not update rar. There are two main uses for this method: 1) To set the initial state, and 2) when using a learned policy, but not updating it.

Here's an example of the API in use:

import QLearner as ql

learner = ql.QLearner(num_states = 100, \ 
    num_actions = 4, \
    alpha = 0.2, \
    gamma = 0.9, \
    rar = 0.98, \
    radr = 0.999, \
    dyna = 0, \
    verbose = False)

s = 99 # our initial state

a = learner.querysetstate(s) # action for state s

s_prime = 5 # the new state we end up in after taking action a in state s

r = 0 # reward for taking action a in state s

next_action = learner.query(s_prime, r)

Part 2: Navigation Problem Test Cases

We will test your Q-Learner with a navigation problem as follows. Note that your Q-Learner does not need to be coded specially for this task. In fact the code doesn't need to know anything about it. The code necessary to test your learner with this navigation task is implemented in testqlearner.py for you. The navigation task takes place in a 10 x 10 grid world. The particular environment is expressed in a CSV file of integers, where the value in each position is interpreted as follows:

  • 0: blank space.
  • 1: an obstacle.
  • 2: the starting location for the robot.
  • 3: the goal location.
  • 5: quicksand.

An example navigation problem (world01.csv) is shown below. Following python conventions, [0,0] is upper left, or northwest corner, [9,9] lower right or southeast corner. Rows are north/south, columns are east/west.

3,0,0,0,0,0,0,0,0,0
0,0,0,0,0,0,0,0,0,0
0,0,0,0,0,0,0,0,0,0
0,0,1,1,1,1,1,0,0,0
0,5,1,0,0,0,1,0,0,0
0,5,1,0,0,0,1,0,0,0
0,0,1,0,0,0,1,0,0,0
0,0,0,0,0,0,0,0,0,0
0,0,0,0,0,0,0,0,0,0
0,0,0,0,2,0,0,0,0,0

In this example the robot starts at the bottom center, and must navigate to the top left. Note that a wall of obstacles blocks its path, and there is some quicksand along the left side. The objective is for the robot to learn how to navigate from the starting location to the goal with the highest total reward. We define the reward for each step as:

  • -1 if the robot moves to an empty or blank space, or attempts to move into a wall
  • -100 if the robot moves to a quicksand space
  • 1 if the robot moves to the goal space

Overall, we will assess the performance of a policy as the median reward it incurs to travel from the start to the goal (higher reward is better). We assess a learner in terms of the reward it converges to over a given number of training epochs (trips from start to goal). Important note: the problem includes random actions. So, for example, if your learner responds with a "move north" action, there is some probability that the robot will actually move in a different direction. For this reason, the "wise" learner develops policies that keep the robot well away from quicksand. We map this problem to a reinforcement learning problem as follows:

  • State: The state is the location of the robot, it is computed (discretized) as: column location * 10 + row location.
  • Actions: There are 4 possible actions, 0: move north, 1: move east, 2: move south, 3: move west.
  • R: The reward is as described above.
  • T: The transition matrix can be inferred from the CSV map and the actions.

Note that R and T are not known by or available to the learner. The code in testqlearner.py will test your code as follows (pseudo code):

Instantiate the learner with the constructor QLearner()
s = initial_location
a = querysetstate(s)
s_prime = new location according to action a
r = -1.0
while not converged:
    a = query(s_prime, r) 
    s_prime = new location according to action a
    if s_prime == goal:
        r = +1
        s_prime = start location
    else if s_prime == quicksand:
        r = -100
    else:
        r = -1

A few things to note about this code: The learner always receives a reward of -1.0 (or -100.0) until it reaches the goal, when it receives a reward of +1.0. As soon as the robot reaches the goal, it is immediately returned to the starting location.

Here are example solutions. Note that these examples were created before we added "quicksand" to the project. We will be updating the examples to reflect this change. In the mean time, you may find these useful:

mc3_p2_examples

mc3_p2_dyna_examples

Part 3: Implement Dyna (5%)

Add additional components to your QLearner class so that multiple "hallucinated" experience tuples are used to update the Q-table for each "real" experience. The addition of this component should speed convergence in terms of the number of calls to query(), not necessarily running time.

Note that it is not important that you implement Dyna exactly as described in the lecture. The key requirement is that your code should somehow hallucinate additional experiences. The precise method you use for discovering those experiences is flexible. We will test your code on several test worlds with 50 epochs and with dyna = 200. Our expectation is that with Dyna, the solution should be much better after 50 epochs than without.

Part 4: Implement author() Method (0%)

You should implement a method called author() that returns your Georgia Tech user ID as a string. This is the ID you use to log into t-square. It is not your 9 digit student number. Here is an example of how you might implement author() within a learner object:

class QLearner(object):
    def author(self):
        return 'tb34' # replace tb34 with your Georgia Tech username.

And here's an example of how it could be called from a testing program:

    # create a learner and train it
    learner = ql.QLearner() # create a QLearner
    print learner.author()

Check the template code for examples. We are adding those to the repo now, but it might not be there if you check right away. Implementing this method correctly does not provide any points, but there will be a penalty for not implementing it.

Contents of Report

There is no report component of this assignment. However, if you would like to impress us with your Machine Learning prowess, you are invited to submit a succinct report.

Hints & resources

This paper by Kaelbling, Littman and Moore, is a good resource for RL in general: http://www.jair.org/media/301/live-301-1562-jair.pdf See Section 4.2 for details on Q-Learning.

There is also a chapter in the Mitchell book on Q-Learning.

For implementing Dyna, you may find the following resources useful:

What to turn in

Turn your project in via t-square. All of your code must be contained within QLearner.py .

  • Your QLearner as QLearner.py
  • Do not submit any other files.

Rubric

Only your QLearner class will be tested.

  • For basic Q-Learning (dyna = 0) we will test your learner against 10 test worlds with 500 epochs in each world. One "epoch" means your robot reaches the goal one time, or after 100000 steps, whichever comes first. Your QLearner retains its state (Q-table), and then we allow it to navigate to the goal again, over and over, 500 times. Each test (500 epochs) should complete in less than 2 seconds. NOTE: an epoch where the robot fails to reach the goal will likely take much longer (in running time), than one that does reach the goal, and is a common reason for failing to complete test cases within the time limit.
  • Benchmark: As a benchmark to compare your solution to, we will run our reference solution in the same world, with 500 epochs. We will take the median reward of our reference across all of those 500 epochs.
  • Your score: For each world we will take the median cost your solution finds across all 500 epochs.
  • For a test to be successful, your learner should find a total reward >= 1.5 x the benchmark. Note that since reward for a single epoch is negative, your solution can be up to 50% worse than the reference solution and still pass.
  • There are 10 test cases, each test case is worth 9.5 points.
  • Here is how we will initialize your QLearner for these test cases:
    learner = ql.QLearner(num_states=100,\
        num_actions = 4, \
        alpha = 0.2, \
        gamma = 0.9, \
        rar = 0.98, \
        radr = 0.999, \
        dyna = 0, \
        verbose=False) #initialize the learner
  • For Dyna-Q, we will set dyna = 200. We will test your learner against world01.csv and world02.csv with 50 epochs. Scoring is similar to the non-dyna case: Each test should complete in less than 10 seconds. For the test to be successful, your learner should find solution with total reward to the goal >= 1.5 x the median reward our reference solution across all 50 epochs. Note that since reward for a single epoch is negative, your solution can be up to 50% worse than the reference solution and still pass. We will check this by taking the median of all 50 runs. Each test case is worth 2.5 points. We will initialize your learner with the following parameter values for these test cases:
    learner = ql.QLearner(num_states=100,\
        num_actions = 4, \
        alpha = 0.2, \
        gamma = 0.9, \
        rar = 0.5, \
        radr = 0.99, \
        dyna = 200, \
        verbose=False) #initialize the learner
  • Is the author() method correctly implemented (-20% if not)

Required, Allowed & Prohibited

Required:

  • Your project must be coded in Python 2.7.x.
  • Your code must run on one of the university-provided computers (e.g. buffet02.cc.gatech.edu).

Allowed:

  • You can develop your code on your personal machine, but it must also run successfully on one of the university provided machines or virtual images.
  • Your code may use standard Python libraries.
  • You may use the NumPy, SciPy, matplotlib and Pandas libraries. Be sure you are using the correct versions.
  • You may reuse sections of code (up to 5 lines) that you collected from other students or the internet.
  • Code provided by the instructor, or allowed by the instructor to be shared.

Prohibited:

  • Any libraries not listed in the "allowed" section above.
  • Any code you did not write yourself (except for the 5 line rule in the "allowed" section).
  • Any Classes (other than Random) that create their own instance variables for later use (e.g., learners like kdtree).
  • Print statements outside "verbose" checks (they significantly slow down auto grading).
  • Any method for reading data besides util.py

Legacy