3. Markov Decision Processes

Your task is to implement the value iteration and policy iteration methods to find the optimal strategy (policy) for the given MDP.

Specification

In the module mdp_agent.py, implement two classes:

  • ValueIterationAgent which will find the optimal strategy using the value iteration method, and
  • PolicyIterationAgent which will find the optimal strategy using the policy iteration method.

The interface of both classes is identical, both must implement the following methods:

method input parameters output parameters explanation
__init__ env: MDPProblem, gamma: float, epsilon: float none Agent initialization.
find_policy none Policy Returns the optimal strategy, i.e., a dictionary of pairs (state, action).
  • The class will be initialized with the following parameters:
    • env is the environment, i.e., an object of type kuimaze2.MDPProblem
    • gamma is the so-called “discount factor” from the range (0,1)
    • epsilon is the maximum allowed error for the values of individual states (used in value iteration)
  • The output of the find_policy() method must be a policy represented as a dictionary, where the key is always a state (instance of the class kuimaze2.State) and the value is the optimal action for that state (instance of the class kuimaze2.Action). The strategy must contain an action for all free states, including terminal ones. The specific action chosen for terminal states does not matter.
  • Timeout for individual runs of value/policy iteration for a given problem instance is set to 30s. (But you should only need a fraction of this time.)
  • The algorithms implemented in the classes ValueIterationAgent and PolicyIterationAgent must correspond to the assignment. For example, it is not allowed to simply call ValueIteration.find_policy() in PolicyIterationAgent.find_policy() or to implement the value iteration algorithm in it (or vice versa). In such a case, the entire task will be evaluated with 0 points!
  • In the implementation of the algorithms, you can only use public methods of the ''MDPProblem'' class. If you feel that you need to use methods of other classes than MDPProblem, or that you need to use non-public variables and methods (whose name starts with _), discuss it with your instructor.

How to

  1. We recommend creating a new working directory for the task. Set up an updated version of the kuimaze2 package in it.
  2. Familiarize yourself with the MDPProblem environment.
  3. In the kuimaze2 package, you will also find the script example_mdp.py, which also shows how to work with the environment. It can be used as a starting code for the implementation of both classes.
  4. It is quite possible that both classes will have some common parts. In such a case, we recommend (as indicated in example_mdp.py) to extract shared parts into a common ancestor of both classes:
    class MDPAgent:
        # Parts common to both methods/agents
        ...
     
    class ValueIterationAgent(MDPAgent):
        # Parts specific for value iteration
        ...
     
    class PolicyIterationAgent(MDPAgent):
        # Parts specific for policy iteration
        ...

Submission

  • The deadline for submitting the task can be found in BRUTE, task 08-MDPs.
  • Submit the module mdp_agent.py, or a ZIP archive with the module mdp_agent.py and other modules you created that your agent needs/imports. These files must be in the root of the archive, the archive must not contain any directories! Do not include/submit any modules that you received from us!

Evaluation

Learn about evaluation and scoring of the task.

courses/be5b33kui/semtasks/03_mdp/start.txt · Last modified: 2024/03/26 20:34 by xposik