Learning_Maze

The Learning Maze is a game to be played by humans. Also, I let an algorithm play it and compare the results from model to human level.

---

Link to web app (Streamlit):

https://luisestrathe-learning-maze-app-kung98.streamlit.app/

Work in progress:

The goal is to have humans and ml algorithms play the same maze. Finally, I will compare the performance of some model to human level. Right now, models are trained and evaluated.

---

FEATURES

• The maze can be configured by size (always quadratic), ratio of walls, number of opponents and how much the agent (player) can see around her.
• The agent can move in the maze up, down, left and right or wait.
• Every map is solvable. The agent can always reach the goal. Yet, opponents might make it impossible to pass.
• scoring will be based on
  • success rate
  • number of steps taken (incl. waiting) in relation to the shortest path
  • the difficulty level
• The difficulty level will be based a regressor trained on the data of the human players. [0 (almost impossible not to win) - 100 (almost impossible to win)]

---

MODEL PERFORMANCE

Human perfornance:

• The human players can play on the web app. Their performance is stored in a Supabase cloud database.

  WIN RATE: 57 %

Model perfornance:

• Agents are acting in random generated maps and see 1 field far

• Performance test over 10 tsd. games each

1. The random baseline model is randomly choosing a move. It is not trained at all. It is just a baseline to compare the performance of the other models to.

   WIN RATE: 2 %

2. The explored q-learning model uses a q-table to choose a move. To not get stuck in a loop, the pvalues additionally are randomly adjusted within a certain range. Nevertheless, the distribution of learned moves remains.

   The pvalues for the moves come from fully random exploration (as in baseline) returning a match score.

   WIN RATE: 46 %

---

REPO STRUCTURE

|── data
│   ├── app             <- images and reports from the web app users
│   ├── input           <- images for maze map generation
│   ├── records         <- records of game results
│   └── tmp             <- temporary files
│
├── src                 <- Source code for use in this project.
│   ├── LM_Run.py       <- Run maze locally
│   └── maze            <- Environment of the maze
│       ├── LM_Data.py  <- Scripts to download or generate data       
│       └── LM_Environment.py
│                       <- Map class and evnrioment functions 
│
│── app.py              <- Streamlit web app
│
├── LICENSE             <- MIT License
├── README   
└── requirements.txt