Does a solution for Wumpus World with neural networks exist?

Question

The Wumpus World proposed in book of Stuart Russel and Peter Norvig, is a game which happens on a 4x4 board and the objective is to grab the gold and avoiding the threats that can kill you. The rules of game are:

• You move just one box for round

• Start in position (1,1), bottom left

• You have a vector of sensors for perceiving the world around you.

• When you are next to another position (including the gold), the vector is 'activated'.

• There is one wumpus (a monster), 2-3 pits (feel free to put more or less) and just one gold pot

• You only have one arrow that flies in a straight line and can kill the wumpus

• Entering the room with a pit, the wumpus or the gold finishes the game

Scoring is as follows: +1000 for grabbing the gold, -1000 for dying to the wumpus, -1 for each step, -10 for shooting an arrow. Fore more details about the rules, chapter 7 of the book explains them.

Well now that game has been explained, the question is: in the book, the solution is demonstrated by logic and searching, does there exist another form to solve that problem with neural networks? If yes, how to do that? What topology to use? What paradigm of learning and algorithms to use?

1*: My English is horrible, if you can send grammar corrections, I'm grateful.

2*: I think this is a bit confusing and a bit complex. if you can help me to clarify better, please do commentary or edit!

Answer 1

Yes! If you read ahead to the chapters in reinforcement learning in the same book, you'll see that the wompus world appears again there. Techniques like Q-learning can be used to solve it, and since Q-learning involves learning the shape of a function, a neural network can be employed as a function approximator.

The basic idea is to treat this problem as an input/output mapping (states -> actions), and to learn which actions produce the greatest rewards.

Note however, that these approaches rely on trial and error. The logic based approach reasons about the rules of the game, and can play reasonably well right away. The learning approach will need to try and fail many times before playing well.