Our new approach aims to allow multiple agent approaches to work together on a wumpus environment. An executable file in java - which can be fed with the locations of the files for two agents - will simulate one walkthrough in our partially observable, strategic, sequential, static, discrete multiagent wumpus world. wumpus world with gym env
based on:
https://github.com/Davood-M/gym_wumpus_world
how to:
https://www.novatec-gmbh.de/en/blog/creating-a-gym-environment/
- install gym
pip install --upgrade gym==0.10.0
- Install wumpus environment into pip
pip install -e .
- execute train.py
python train.py
- wumpus environment is rectangular
- Agents: one or more wumpi, two Player
- possible observations: nothing, gold, pit, breeze, stench, exit (mby light), locationvector for scream
- gold: contains one gold
- pit: kills agent
- breeze: in von Neumnann neighborhood to pit
- stench: in von Neumnann neighborhood to wumpus
- vector: pointing from one agents position to the scream location
- Rectangle: minimal size: |players|+2 for hight or width, the other side being at least |players|
- all gold has to be reachable
- ...pits: |players| + 1
- ...gold: |wumpi|
- ...wumpi: |gold|
Perception of a player is limited to the cell she is standing on (partially-observable) This means the state of only this one cell is fixed but she is meant to remember the states of cells she has been to since they dont change a lot.
- walk_up
- the player can walk one cell every move
- walk_down
- the player can walk one cell every move
- walk_left
- the player can walk one cell every move
- walk_right
- the player can walk one cell every move
- pick up
- the player can pick up gold if she is on a gold cell
- put down
- if the player has gold she can put it back on the ground
- climb
- the wumpus cave has to be left climbing out the way the player entered
- scream
- a scream can be heard by other players and will be represented as a vector
- nothing
- do nothing for one turn
pick up a piece of gold and leave (asap?)