Simulator for the AI Wumpus World game in C++ with Python wrapper.
Written by Larry Holder ([email protected]).
The Wumpus Simulator is a simple framework for simulating the Wumpus World described in Russell and Norvig's "Artificial Intelligence: A Modern Approach". The simulator is written in C++, but you can design your agent in either C++ or Python. The idea is for you to modify the Agent.h and Agent.cc (or Agent.py) files to implement your super-smart agent. The C++ and Python agents provided simply accept commands from the keyboard and execute them in the simulator. The keyboard commands are 'f' for forward, 'l' for turnleft, 'r' for turnright, 'g' for grab, 's' for shoot, and 'c' for climb.
To try out the simulator, install the code on a UNIX system (or a system that has the 'make' program installed and a C++ compiler). Type 'make' to build the 'wumpsim' executable. Then, type './wumpsim'. You should see a randomly-generated 4x4 world, information about the game state, and a prompt for the next action. When the game is over, scoring information is provided.
If you want to code your agent in Python, see the "Python Agent" section below.
The wumpus simulator takes a few options, as described below.
-size <N> lets you to set the size of the world to NxN (N>1). Default is 4.
-trials <N> runs the simulator for N trials, where each trial generates a new
wumpus world. Default is 1.
-tries <N> runs each trial of the simulator N times, giving your agent
multiple tries at the same world. Default is 1.
-seed <N> lets you set the random number seed to some positive integer so that
the simulator generates the same sequence of worlds each run. By default, the
random number seed is set semi-randomly based on the clock.
-world <file> lets you play a specific world as specified in the given file.
The -size option is ignored, and each try and trial uses the same world. The
format of the world file is as follows (all lowercase, must appear in this
order):
size N
wumpus N N
gold N N
pit N N
pit N N
...
where N is a positive integer. Some error checking is performed. A sample world file is provided in testworld.txt.
The simulator works by generating a new world and a new agent for each trial. Before each try on this world, the agent's Initialize() method is called, which you can use to perform any pre-game preparation. Then, the game starts. The agent's Process() method is called with the current Percept, and the agent should return an action, which is performed in the simulator. This continues until the game is over (agent dies or leaves cave) or the maximum number of moves (1000) is exceeded. When the game is over, the Agent's GameOver() method is called with the final score. If additional tries are left for this world, then the world is re-initialized, and the agent's Initialize() method is called again, and play proceeds on another instance of the same game.
After the number of tries is completed, the agent is deleted. So, you may want to store some information in the agent's destructor method to be reloaded during the agent's constructor method when reborn for the next trial. If additional trials have been requested, then a new wumpus world is generated, and the process continues as described above.
Scoring information is output at the end of each try, each trial and at the end of the simulation.
Your agent must include at least five methods: constructor, destructor, Initialize, Process, and GameOver. You may change any or all of these methods to implement your agent. And you may include additional methods as you see fit. You may also take advantage of the other classes defined for the simulator (e.g., Percept, Action, Orientation, Location, WorldState, WumpusWorld) by including them directly, subclassing them, or borrowing code for your agent. Once your agent is complete, simply type 'make' to remake the simulator with your C++ agent. Then run 'wumpsim' to try it out. Or, for a Python agent, simply run 'pywumpsim'; no need to recompile the simulator.
You can also implement your agent in Python 2 or 3. First, you will need to compile a different version of the simulator that supports calling external Python agent functions. Instead of 'make' or 'make wumpsim', you should instead compile the simulator using 'make pywumpsim'. On some platforms you need to set the PYTHONPATH environment variable to your working directory, where pywumpsim and Agent.py reside.
Next, you will make all your changes to the Agent.py file. This file contains the Agent class consisting five methods: constructor (__init__), destructor (__del__), Initialize, Process, and GameOver. Note that the Process method takes a percept class as defined in Percept.py. The Process function should return one of the six actions defined in Action.py.
Once you've finished your Agent.py file, simply run the 'pywumpsim' program to test your agent. The Agent.py file and the 'pywumpsim' executable must be in the same directory. The 'pywumpsim' program accepts all the options described above for 'wumpsim'. And there is no need to recompile 'pywumpsim' after changes to Agent.py; this file is reloaded everytime 'pywumpsim' is executed.
Happy hunting!
Thanks to Sergio Tessaris for updates to support Python 3.