What is the Wumpus World?
Making it out alive from this common AI / algorithm environment
Many students are becoming familiar with the Wumpus World through the college textbook Artificial Intelligence A Modern Approach. The idea for the world came from the 1973 video game “Hunt the Wumpus” by Gregory Yob¹ and has since been used by many to examine the effects of modeling different kinds of algorithms.
The purpose of the world is to provide a simple goal for the model to solve. This way, the user can focus on the implementation of the algorithm, not on having to model and account for endless conditions. So, the Wumpus World is a tool to help an individual learn how an algorithm works.
The Wumpus World is not unique; there are several example “worlds” that can be used to model to different AI approaches. For example, Vacuum World and Block World. Each have different laws, or rules, regarding how to interact with them.
In this world, you can image yourself as the agent. This is the character that is navigating and exploring with the goal of collecting gold and leaving the cave without getting eaten by the wumpus or falling down any pits.
To the agent, the world is partially-observable. This means that the agent can sense things in cells directly around them (current, up, left, right, and down), but not anywhere else. So, if your agent is at the start in cell (1,1), it will not know if there is a stench in cell (1,4).
An agent can try to avoid dying (getting eaten or falling) and know when to pick up gold by responding to the perceptions around it.
😨 Stench: If the agent is in a cell directly next to the wumpus (not diagonal, the agent can only move up and down and left and right), it will perceive a stench. This is a warning to the agent that one of the cells around it contains the wumpus. If the agent enters the wumpus’ room accidentally, the agent will be eaten.
😱 Breeze: If the agent is next to a pit, they will perceive a breeze. This is a warning that on the agent’s next move, if they choose wrong, they will stumble into a pit and fall to their death.
😌 Glitter: If the agent is in the room containing gold, the agent will perceive glittering. This is a great sign and notification that the agent should pick up the gold.
☠️ Scream: If the agent decides to hunt down the wumpus and shoot it, it will emit a scream to let you know that it has died. A bit morbid, but it was going to eat you! This is a good sign as it means your agent no longer has to worry about accidentally running into the wumpus.
Based off the perceptions an agent has while in a cell, they can perform actions. What action is chosen is often based on what type of algorithm you have selected to implement. Different algorithms will propose different actions. Possible actions are
🏃♀ Move: An agent can decide to move up, down, left, or right to the cell they are currently in.
🏹 Shoot: If the agent perceives a stench, and they think they know what cell contains the wumpus, they can choose to shoot into a cell, hoping to kill it and have it emit a scream. An agent only has one arrow to shoot with however, so they must be certain!
👑 Pick up Gold: When the cell the agent is in is glittering, they should pick up the gold on the ground.
☝️ Climb: Once the agent has collected all the gold they can, they can escape the cave by climbing out, which will end the game.
And that’s about it! The world can sometimes be modified to make it simpler (not possible to shoot the wumpus) or more complex (every action has an associated cost), but these are the bare basics.
If you’d like to check out a simulation of the Wumpus World, as shown in the screenshot above, you can visit here and play the game. It’s a little difficult to see the perceptions while you’re in a square, but I got used to moving back a square whenever I saw that the perceptions were written behind the agent.
Look out for tutorials soon from me around how different algorithms work in the Wumpus World. For now, here’s my other Wumpus World readings!
