Wumpus world

The Wumpus world is a simple environment for knowledge-based software agents that is often used in teaching to explain the logical functioning of software agents. The environment comes from the computer game Hunt the Wumpus .

Modeling and task environment

Example of a 4 * 4 field Wumpus world

The Wumpus world represents a cave and is modeled by a square grid. At each point in time and for each field of the grid it is clearly defined what or who is in this field or whether the field is empty. The wumpus is hiding in a field. There is gold on another space. The remaining fields either have pitfalls or are empty. These fields are pre-determined and do not change.

The task of the software agent is to find the gold in the shortest possible time and return to the starting position unharmed. The figure of Wumpus and several trap doors pose a danger. Initially, the agent has no knowledge of what is where in the cave. With the help of five sensors , however, he is able to perceive his immediate surroundings and with the help of these perceptions gain new knowledge about the cave through logical conclusions . On the basis of this knowledge, he makes the decision where to move next in order to perceive his surroundings again.

The agent starts on a defined field and moves in each step to an adjacent field (shared edge with the field on which it is located). With the help of a sensor, the agent is able to perceive a trap door that is located on an adjacent field (draft). However, he cannot perceive which of the adjacent fields this is. It is the same with the perception of the wumpus (stench). The gold is only recognized when the agent is directly on this field (glitter). A fourth sensor is used to sense a shock when the agent hits a wall. The agent also has exactly one arrow with which he can shoot in one direction (horizontally or vertically) to the nearest wall or to the wumpus. The fifth sensor is used to perceive the scream that sounds when the Wumpus is hit by the arrow, which means its death.

A precise definition for the agent's task environment can be given with the PEAS description:

Performance: Each action of the agent is rated with a certain number of points. Picking up the gold: +1000, entering a space with a trapdoor or the wumpus space if the wumpus is still alive: -1000. Using the arrow: -10. Any action: -1.
Environment: The above-described grid with gold, trap doors and wumpus.
Actuator: The actions that the agent can take. Moving forward, turning 90 degrees to the left or right, picking up the gold, shooting the arrow.
Sensors: The five sensors described above for sensing the environment.

example

The rows and columns of the grid in the above figure are numbered from 1 to 4, the rows are counted from bottom to top, the columns from left to right. Each field is labeled with , where the row number and the column number denote. The agent's starting field is . ${\ displaystyle (x, y)}$ ${\ displaystyle x}$ ${\ displaystyle y}$ ${\ displaystyle (1,1)}$

The agent on the starting field has no knowledge. Not feeling a draft or a smell, he concludes that the adjacent fields to the right and above are safe. Depending on the strategy, he now decides on one of the two fields, in this example he goes to the right. Here he perceives the draft and concludes that there is a trap to the right of or above him. It is not yet possible to determine the exact position of the trap. ${\ displaystyle (1,2)}$ ${\ displaystyle (2,1)}$

The agent now has the following knowledge about his environment :: sure. There is a trap door on or (the “or” is not exclusive, it can also be a trap door on both spaces). On the other hand, there is neither on nor on the Wumpus, as he did not perceive any stench in an adjacent field. ${\ displaystyle (1,1), (1,2), (2,1)}$ ${\ displaystyle (1,3)}$ ${\ displaystyle (2.2)}$ ${\ displaystyle (1,3)}$ ${\ displaystyle (2.2)}$

In this case, the agent goes into the only safe area that he has not yet entered. Here he perceives a stench, from which he deduces that a wumpus is nearby. However, he has no draft, so he concludes that there cannot be a trapdoor to his right or above . Since he already knows that there is no Wumpus, he can now identify this field as safe and the agent decides to enter this field. Alternatively, he could have fired his arrow at the wumpus, since with his knowledge of Feld he could safely conclude that the wumpus must be on. On he takes nothing true and may his decision from the two safe boxes (about him) or take (to his right). After the first decision he finds the gold immediately, after the second he perceives a draft and can therefore not deduce any further safe field. The only safe field remains for him . So he goes back on and makes the safe move on the gold. ${\ displaystyle (2,1)}$ ${\ displaystyle (2.2)}$ ${\ displaystyle (3,1)}$ ${\ displaystyle (2.2)}$ ${\ displaystyle (2.2)}$ ${\ displaystyle (3,1)}$ ${\ displaystyle (2.2)}$ ${\ displaystyle (3.2)}$ ${\ displaystyle (2,3)}$ ${\ displaystyle (3.2)}$ ${\ displaystyle (2.2)}$

literature

Stuart Russell, Peter Norvig: Artificial Intelligence Pearson Verlag, Munich 2004, ISBN 3-8273-7089-2

Web links

Stuart Russell, Peter Norvig: Logical Agents . Chapter 7 of the textbook, Artificial Intelligence: A Modern Approach (Second Edition) (PDF file; 417 kB)