Evolving neural networks-controlled virtual pets -- a petri dish in your computer.
I'm very faschinated by evolutionary algritms, virtual life and artificial intelligence. This project combines them, and after I had the iea, it has been irresistable to me.
NeuPet is heavily inspired by a MS-DOS game called Evolve! Lite I found on a shareware CD in the nineties, called Game Empire. It used pseudo celluar automata to manage a population of "creatures" moving around on a grid, acording to a configurable set of rules. The creatures would respond to the prescense of friends or foes (hardcoded as two spicies) and go forward, left, right, etc.
A set of "pets" live in a hexagonal grid world. Each pet is controlled by a neural network, and has to find food and avoid getting eaten in order to survive and reproduce.
This is not just an evolutionary algorithm, since there is no concept of "generations" or any sort of fitness evaluation. pets are just let loose in the world and the ones who survive the longes, and have the most offspring get to pass on their genes. Just like in real life.
- Interface for the Pet position and direction, energy levels etc. in World.
- Split the plant logic to separate class with generic base class. (?) Not fun.
- Regenrerate the entire world after extinction. (random walls etc?)
- Make pets sense their own energy level, age etc.
- Add hazardous areas. Water? Lava?
- Health, Strength, Metabolism, Mass etc. interacting with each other. Dead Pets should still contain some energy to be useful as food.
- Pet "memory". Implement as an output feeding the input value of the next step.
- New senses.
- Sight (1/distance to closest Pet) in n directions over m degrees of FOV. Multiple channels for food, mate, wall and danger?
- Smell - Each world cell should have the smell of the last occupant. Outputs to the visiting Pet is the strenght of the smell, and the gene similarity to itself. (?)
- Bundle the Pet and PetState into a common stuct for storage in World. Makes the std::map from Pet to PetState unnecesary.
- Decompose dead pets slowly.
- Decomposing Pets should seep into the ground, adding nutricients (pland growth rate) for plants.
- Plants should deplete the nutricients of the ground.
- Pooping Pets add fertilizer to the ground?
- Water (with Pets able to traverse water or land by varying speed? Value for how land or water-bound they are? Some Pets amphibious. How to ballance it?)
- WorldCell height - (Replaces impassable walls and water (water level = 0).) Higher energy consumption uphill, impossible to climb too steep hills (gradient > 1), dangerous drops. (Needs 3D (or 2.5D) graphics.)
- Speciation. Add an input for how many genes a pair of Pets share.
- Save & load Pet population. JSON?
- Save and load World layout.
- Cycle through several World layouts with a set interval. Enbles training Pets for moving fields of plants etc.
- Non-neural Pet class until I get the world and evolution working.
- GUI. Picking and examining Pets, etc.
- Ancestry tracking. Useful for statistics.
- Pretty graphics. http://www.youtube.com/watch?v=-PsQqpK0MvM&list=UUutMBmI_ydvgmzbkQxbI4IQ&index=9&feature=plcp
This will be impotant to make the simulation sensislbe. They are the laws of physics in the World.
Genetically controlled values
- birthIdealMass - Mass of Pet at birth.
- targetIdealMass - Mass of fully grown Pet.
Simulation controlled values
- energyConversionFactor = 10
- breedEnergy - birthIdealMass * childrenPerBirth * energyConversionFactor
- idealMass -
- maxMass = Target+50%, minMass )
- mass - (Mass gained by eating = mass of food / energyConversionFactor)
- Energy content/nutriciency => directly proportional to Mass.
- Greater mass => longer pregnancy (directly proportional, plus constant) http://www.applet-magic.com/gestation.htm
- Greater mass => greater strength (mass^1/3 proportional to strength^1/2 ?)