leshchenko1979 / fishbowl Goto Github PK

• init
• activation

It would be nice to give the fish a sense of how far the food is. It would certainly give them hint how big the pulse should be.

Probably use https://github.com/wagenaartje/neataptic/tree/master/graph - should work well with carrot's networks.

Consider introducing disparity in size decrease based on:

1. disparity in initial sizes.
2. disparity in speed.
3. disparity in hunger (eatenTimer).

Should also fix #39

So that further mutations are standing on the shoulders of giants instead of starting from scratch each time the app is launched

was broken by #34

• decide on generations design
  • for now mutations only on respawn
  • splitted fish refers to the same brain in the neat. population array as its parent
• fish respawn after X frames
• implement a fitness function
  • the total size of all fish sharing this brain after X frames + max witnessed size of all fish sharing this brain
• a block with mutation parameters
• handle mutation on split
• handle mutation on respawn after all fish die

Needs fixing the bounce mechanics

All fish start from one point in the direction of one food
Food is positioned randomly in the area where all the fish are directed
Fitness is 1 / closest dist to the food
Shorter gen duration (approx the time it takes the fish to reach the food directly plus some leeway)
Sim running until all fish in the generation hit the food

1. The fitness function is how long a certain fish has survived.
2. One of the fish (or both of them) mutates on split.
3. When the simulation time elapses for a generation, surviving fish are evaluated on their lifetime.
4. When the simulation ends because all of the fish have died out, the score is proportional on the recorded lifetime of the corresponding fish.

• start recording birthFrame in Critter() constructor
• on split clone brain and add it to neat
• mutate fish on split
• reset birthFrame on split?
• when a fish dies, set its brain's score to the fish lifetime
• calculate neat.population[i].score based on lifetime (currentFrame - birthFrame)
• entirely replace neat.population on generation change instead of neat.mutate()

Originally posted by @leshchenko1979 in #20 (comment)

• generation no. / frame no.
• max genome size
• current max fitness
• NN response distribution

It’s time to implement vision and turn steering actions back on again.

Food perception can be implemented effectively through reading the foodDensity array instead of trying to analyze the position of every food piece.

Other fish can be sensed just by going through the fish array and incrementing a counter by 1/distance.

• fish perception
• food perception
• turn on turning actions
• introduce acceleration???

Should probably try neat.evolve () instead of producing offspring one by one

Replace with per-generation stats or decimate data

Make it a 2D-array - will help avoid bugs

Investigate and, if confirmed, insert an additional network into neat.population when splitting the fish

Evolution is stifled because immobile fish that grow at the start due to appearing on top of food get promoted unfairly

Fish quickly and steadily lose weight and die

Food density as a way to discover different evolution strategies doesn't serve its role.

Food density in the eyes of fish is proportional to the (amount of food / area of the fishbowl / fish max speed). Secondly, although the starting amount of food may be high, the amount of other fish may quickly drive it down.

So, the "food amount parameter" should be replaced with food density at least, or, maybe, with food density / fish density.

Their sum should be 100, but it isn't

Use https://getbootstrap.com/docs/4.4/components/modal/#vertically-centered

On smaller screens you need to scroll to get to the options. This should be corrected by moving some of the contents of the left column to a pop-up / modal / accordeon.

• limit pulse when speed is too high
• limit split if the size is too small
• channel results of thinking through a variable in fish
• introduce turn left / turn right actions

the food on the right side of the screen stops growing while the left goes on

Give the user the ability to turn off / on senses and actions

E.g., if the result of brain.activate is between 0 and 1, then accelerate in between 0% and 100% accordingly.

Same with turning.

Conical vision makes it easier to gradually zoom in on the target as you move. Linear vision (like now) makes the fish lose a moving target and miss a lot of food that's away from the view axis.

Leave only speed as input and split and pulse as outputs