Revision f9c10f9fb250601bee5825b36a8c8c30d66e2eef (click the page title to view the current version)

MDP

Changes from f9c10f9fb250601bee5825b36a8c8c30d66e2eef to 54c78ad01d667d5ac7a191819eb4ab58f5eaa0a3

---
title: Markov Decision Processes
categories: session
---

This session is a precursor to reinforcement learning.
This session is a precursor to 
[Reinforcement Learning]().

+ **Reading:** Russel and Norvig, Chapter 16.1 and Chapter 23.1
+ **Reading:** Russel and Norvig, Chapter 16.1-3 + cursory 16.4
+ [A video demo](https://s3.amazonaws.com/media-p.slid.es/videos/1961471/NL0eUKU3/learning_dexterity.mp4)
+ [Briefing on MDP]()
+ Eirik's lecture notes from 2022
    + [Slides (PDF)](Reinforcement Learning Slides 1.pdf)
    + [Notes (PDF)](Reinforcement Learning Notes 1.pdf)

# Exercises

## Task 1

Recall the requirements to model a problem with the MDP framework:

- Sequential decision problem     
- Stochastic Environment
- Fully observable
- with a Markovian Transition model
- with additive rewards

And the properties:
+ Recall the requirements to model a problem with the MDP framework:
    - Sequential decision problem     
    - Stochastic Environment
    - Fully observable
    - Markovian Transition model 
        - the probability distribution depends only on the current state and not past history
    - Additive rewards
+ and recoll the model elements
    - A set of states $s \in S$
    - A set of actions $a \in A$
    - A probabilistic transition function $T(s,a,s')$
    - A reward function $R(s,a,s')$
    - A start state $S_0$
+ Consider one of the following problems
    - The Frozen Lake
    - The game of Black Jack
    - any problem you have previously studied in the semester, e.g. on CodinGame
    - The moon lander (see the [animation from Gymnasium](https://gymnasium.farama.org/))
+ Form groups of 2-4 students and choose one problem each from the
  list above.

- A set of states $s \in S$
- A set of actions $a \in A$
- A transition function $T(s,a,s')$
- A reward function $R(s,a,s')$
- A start state $S_0$
### Individual Part

### Part A
Analyse the chosen probler, by

Find a problem on CodinGame (preferably one you have worked on), and check if it fulfills the requirements above.
If not, can you think of how you can change the problem (e.g. by adding randomness to actions)? If this is not possible, try with another problem.
+ checking if it fulfills the requirements above
+ identify all the model elements (above)

### Part B
If the problem does not satisfy the MDP requirements, 
can you amend it so that it does?
E.g. could the game be changed into a random one?
If it cannot be adapte, try another problem.

Using the properties of an MDP:

Can you make a graphical representation of the modified problem from Task A? You can chose a subset of the states (and actions if necessary) to reduce the size of the representation. Use either the Dynamic Decision Network from the book (ch 16.1.4), or a simple representation as was done on the slides. (E.g. from [here](https://towardsdatascience.com/real-world-applications-of-markov-decision-process-mdp-a39685546026) or [here](https://en.wikipedia.org/wiki/Markov_decision_process#/media/File:Markov_Decision_Process.svg) )

### Part C
+ Make a graphical representation of the (possibly modified) problem from Task A.
    - if the state space is very large, you may draw only a subset for illustration
    - you can use either the Dynamic Decision Network (R&N ch 16.1.4), a state machine
      graph, or any other simple representation (e.g. like 
      [this](https://towardsdatascience.com/real-world-applications-of-markov-decision-process-mdp-a39685546026)
      or [this](https://en.wikipedia.org/wiki/Markov_decision_process#/media/File:Markov_Decision_Process.svg) )

- Does the problem have a finite or infinite horizon? 
- If you were to attempt to solve the MDP, could the current horizon pose a problem, why/why not?
### Group Work

### Part D
1.  Present your drawings and analysis to each other
    - is the analysis convincing?
2.  Discuss together,

- Does the problem have a discounted reward?
- If you were to attempt to solve the MDP, what discount factor would make sense to use for the utility function?
    - Does the problem have a finite or infinite horizon? 
        - If you were to attempt to solve the MDP, could the current horizon pose a problem, why/why not?
    - Does the problem have a discounted reward?
        - If you were to attempt to solve the MDP,
          what discount factor would make sense to use for the utility function?

## Exploring the Frozen Lake (Task 2)

We will be using the Gymnasium framework to test concepts and ideas
from reinforcement learning.  You may want to consult the documentation,
but you should try playing with it first.

+ [Gymnasium](https://gymnasium.farama.org/)
  formerly known as Gym from OpenAI
+ [Frozen Lake](https://gymnasium.farama.org/environments/toy_text/frozen_lake/)

Install with these statements.  (I am not sure if you need pygame or not.)
```bash
pip install gymnasium
pip install pygame
```


### Part A

**Familiarize yourself with the FrozenLake environment**
You can import and start the simulation of the Frozen Lake like so:
```python
import gymnasium as gym
env = gym.make('FrozenLake-v1', desc=None, map_name="4x4", is_slippery=True,render_mode="human")
obs, info = env.reset(return_info=True)
```
You should now hopefully see a render of the environment.

**Try out some of these functions and see what they do:**
```python
env.action_space.sample()
observation, reward, terminated, truncated, info = env.step(env.action_space.sample())
env.P # The MDP
env.P[s] # Transition matrix of state s
env.P[s][a] # Transitions from state s given action a
```

**Try to make a custom Frozen Lake map**

When creating a frozen-lake environment you can add a custom-map with the ```desc``` argument, e.g:
```python
fl_map = ["HFFFS", "FHHFF", "FFFFH", "HFFHG"] 
env = gym.make('FrozenLake-v1', desc=fl_map, is_slippery=True)
```
For a custom 5x4 map.

## Task 3

Recall the value/utility-function:

$$U(s) = \mathop\max\limits_{a \in A(s)} \sum\limits_{s'}P(s'|s,a)[R(s,a,s') + \gamma U(s')]$$

The Q-Function:

$$Q(s,a) = \sum\limits_{s'}P(s'|s,a)[R(s,a,s') + \gamma \mathop\max\limits_{a'}Q(s',a')]$$

And the function to extract an optimal policy from the Q-Function:

$$\pi^*(s) = \mathop{\mathrm{argmax}}\limits_aQ(s,a)$$

### Part A

Implement the above functions in Python

### Part B

Given a FrozenLake map, and a list of pre-calculated expected utilities,
(e.g.:
```python
utilities = [0.41,0.38,0.35,0.34,0.43,0,0.12,0,0.45,0.48,0.43,0,0,0.59,0.71,1]
```
for the default FrozenLake 4x4 map)

- Test out the utility-function, and see if it makes sense/work as it should.
- Use the function to extract an optimal policy to move around the map (discount factor can be e.g. 0.99).