The Value and Worth of a Food Sensor
45 views
Skip to first unread message
Julian Hershey
May 26, 2016, 1:31:53 AM5/26/16
to Scriptbots
Hi all!
I wanted to make a quick discussion post to maybe drum up some talk and interest in Scriptbots, since I think I have made a lot of headway in the last few months making my own stable version.
Specifically I want to talk about something I actually removed from my version many, many months ago: the food sensor.
At the time, I eliminated it because I felt it was too easy for bots to know the amount of food underneath, as they would just stop and eat it until it was gone. The game of figuring out where the food was located was simple and easy. So I tried letting agents see food cells. The project was a massive failure. Not only was processing speed reduced, but agents never really figured out that green things were food and red things were meat. I commented out the code. Next I tried food "smell", where the number of cells occupied by food is counted, but so far I haven't seen anything that uses this very efficiently. So right now, the typcial way my agents find food is they don't. Now they just use eyesight or hearing "pressure" to drive away from other bots, increasing their chances of running over food.
I wonder what the best "solution" for this is. When I think of ways animals know where the food is in real life, I think of eyesight, smell, inborn habit (like with newborn mammals), and even touch. So should each of these senses have code that cell layers can effect? Can another way be explored? Or should I return to the old food sense way, having over-thought this aspect of the sim? I welcome any input. =)
I wanted to make a quick discussion post to maybe drum up some talk and interest in Scriptbots, since I think I have made a lot of headway in the last few months making my own stable version.
Specifically I want to talk about something I actually removed from my version many, many months ago: the food sensor.
At the time, I eliminated it because I felt it was too easy for bots to know the amount of food underneath, as they would just stop and eat it until it was gone. The game of figuring out where the food was located was simple and easy. So I tried letting agents see food cells. The project was a massive failure. Not only was processing speed reduced, but agents never really figured out that green things were food and red things were meat. I commented out the code. Next I tried food "smell", where the number of cells occupied by food is counted, but so far I haven't seen anything that uses this very efficiently. So right now, the typcial way my agents find food is they don't. Now they just use eyesight or hearing "pressure" to drive away from other bots, increasing their chances of running over food.
I wonder what the best "solution" for this is. When I think of ways animals know where the food is in real life, I think of eyesight, smell, inborn habit (like with newborn mammals), and even touch. So should each of these senses have code that cell layers can effect? Can another way be explored? Or should I return to the old food sense way, having over-thought this aspect of the sim? I welcome any input. =)
Samuel Mungy
Jun 4, 2016, 1:25:45 PM6/4/16
to Scriptbots
In an program of mine which is like scriptbots, I have the organisms see the food and the results are varying. Sometimes, organisms move around looking for food but with limited overall success. The most succesful organisms tend to be the ones who go around randomly and come across food more often. Which is certainly weird. Maybe by adding physical barriers in the world would help by which evolution would actively select organisms who can overcome these barriers and find food.
On a side not, I wanted to know whether the bots are penalized for overeating?
Julian Hershey
Jul 7, 2016, 12:41:21 AM7/7/16
to Scriptbots
Hi Samuel,
Barriers would be helpful (ironically), but then a new question: should the bots be able to see the barriers, and if so, how? In most Neural Network program problems, there is a clear goal, and scriptbots just doesn't have that. Then again, life doesn't have that, it just happens that whatever reproduces the most, wins...
To answer your question, they are not. Andrej had a (very) old version in which the agents were meant to optimize towards a value of 1 in the range [0,2], and would loose fitness if their health got too great. I don't agree with that raw approach, as it's too clear what the solution is. I would think making overeating agents slower and lose reproduction value would be better. I'll consider adding those features in tandem with bringing food-sense back, and see what happens. Could have a single clear solution again, however. We'll see.
=J
Barriers would be helpful (ironically), but then a new question: should the bots be able to see the barriers, and if so, how? In most Neural Network program problems, there is a clear goal, and scriptbots just doesn't have that. Then again, life doesn't have that, it just happens that whatever reproduces the most, wins...
To answer your question, they are not. Andrej had a (very) old version in which the agents were meant to optimize towards a value of 1 in the range [0,2], and would loose fitness if their health got too great. I don't agree with that raw approach, as it's too clear what the solution is. I would think making overeating agents slower and lose reproduction value would be better. I'll consider adding those features in tandem with bringing food-sense back, and see what happens. Could have a single clear solution again, however. We'll see.
=J
Reply all
Reply to author
Forward
0 new messages