Home Download Help Forum Resources Extensions FAQ NetLogo Publications Contact Us Donate Models: Library Community Modeling Commons Beginners Interactive NetLogo Dictionary (BIND) NetLogo Dictionary User Manuals: Web Printable Chinese Czech Farsi / Persian Japanese Spanish

NetLogo Models Library: 
If you download the NetLogo application, this model is included. You can also Try running it in NetLogo Web 
Robby the Robot is a virtual robot who moves around a room and picks up cans. This model demonstrates the use of a genetic algorithm (GA) to evolve control strategies for Robby. The GA starts with randomly generated strategies and then uses evolution to improve them.
Robby's 10x10 square world contains randomly scattered cans. His goal is to pick up as many as he can. At each time tick, Robby can perform one of seven actions: move in one of the four cardinal directions, move in a random direction, pick up a can, or stay put.
When Robby picks up a can, he gets a reward. If he tries to pick up a can where none exists, or bumps into a wall, he is penalized. His score at the end of a run is the sum of these rewards and penalties. The higher his score, the better he did.
To decide which action to perform, Robby senses his surroundings. He can see the contents of the square he is in and the four neighboring squares. Each square can contain a wall, a can, or neither. That means his sensors can be in one of 3<sup>5</sup> = 243 possible combinations.
A "strategy" for Robby specifies one of his seven possible actions for each of those 243 possible situations he can find himself in.
(Advanced note: If you actually do the math, you'll realize that some of those 243 situations turn out to be "impossible", e.g., Robby will never actually find himself in a situation in which all cardinal directions contain walls. This is no problem; the genetic algorithm essentially ignores the "impossible" situations since Robby never encounters them.)
There are many possible variations on the basic concept of a genetic algorithm. Here is the particular variant implemented in this model.
We begin with a pool of randomly generated strategies. We load each strategy into Robby in turn, and then run that strategy in a series of randomly generated arrangements of cans ("environments"). We score Robby on how well he does in each environment. If Robby hits a wall, he loses 5 points. If he successfully picks up a can, he gains 10 points. If he tries to pick up a can, but there isn't one there, he loses 1 point. Robby's average score across all these environments is taken as the "fitness" of that strategy.
Once we have measured the fitness of the current pool of strategies, we construct the next generation of strategies. Each new strategy has two parents. We pick the first parent by picking 15 random candidate parents, then choose the one with the highest fitness. We repeat this to pick the second parent. Each pair of parents creates two new children via crossover and mutation (see below). We keep repeating this process until we have enough new children to fill up the population (settable by the POPULATIONSIZE slider). This "generation" of new children replaces the previous generation. We then similarly calculate the fitness of the current generation, choose parents, and create a new generation. This process continues as long as the GO button is pressed, or is controlled by the NUMBEROFGENERATIONS slider when the GONGENERATIONS button is pressed.
To combine the strategies of two parents, we use "crossover". We pick a random crossover point and combine the first part of the first parent's strategy with the second part of the other's parent's strategy. For example, if the crossover point selected is 50, we use the first 50 entries of the first parent's strategy and the last 193 entries of the second parent's strategy. (Each strategy is in a fixed order.)
In addition to crossover, the children's strategies are subject to occasional random mutation (settable by the MUTATIONRATE slider), in which an action is replaced by a randomly chosen action.
Press SETUP to create the initial pool of random strategies. Press GOFOREVER to start the genetic algorithm running, or GONGENERATIONS to run the genetic algorithm for a fixed number of generations (settable by the NUMBEROFGENERATIONS slider).
In the view, you'll see the pool of strategies (represented by "person" icons). The strategies are heterogeneous. The diversity in their fitness is visualized by color and position on the xaxis. Their color is a shade of red, scaled by their fitness. It's lightest when the fitness is low and darkest when it's high. In addition, the fitness of a strategy determines where it is along the xaxis, with the least fit strategies on the left, and the fittest strategies on the right.
Another aspect of diversity is the difference between the strategies. This difference is measured by counting the frequency of each of the 7 basic actions in the strategy (the "allelediversity"), forming a 7 dimensional vector and then calculating the Euclidean distance between two such vectors. One of the strategies with the highest fitness is placed in the center of the yaxis. The other strategies are placed at locations whose distance from the center along the yaxis is proportional to the difference between their strategy and the winning strategy.
All of this takes a fair amount of time to show, so for long runs of the GA you'll want to move the speed slider to the right or uncheck the "view updates" checkbox to get results faster.
Any time you want to pause the algorithm and see how the current best strategy behaves, press GOFOREVER and wait for the current generation to finish. Recheck "view updates" if you unchecked it before. Next, press VIEWROBBY'SENVIRONMENT. This displays the grid that Robby moves in, with a new, random distribution of cans. Then, press STEPTHRUBESTSTRATEGY. Each time you press this button, Robby will use the best strategy from the last generation to take an action in the current environment. Keep pressing the STEPTHRUBESTSTRATEGY button to see how the strategy works. At any time you can press VIEWROBBY'SENVIRONMENT again to start over with a new environment of cans. If you want to go back to the "strategy view", press the VIEWSTRATEGIES button.
Robby's performance gradually improves. How long does it take to get to a medium or high fitness?
How does Robby typically behave with a totally random strategy? How does he behave once some evolution has taken place?
Does Robby's performance eventually reach a plateau? How does he behave with the strategies that ultimately evolve?
On the way to the final plateau, were there ever any temporary plateaus?
Vary the settings on the POPULATIONSIZE and MUTATIONRATE sliders. How do these affect the best fitness in the population as well as the speed of evolution?
Add a slider for "crossover rate", the probability that two parents create offspring by crossover. If they don't crossover, they simply clone themselves (and then the cloned offspring undergo possible mutation). (You'll need to change the crossover
procedure.)
Add a slider for the tournamentsize. How does varying the tournamentsize affect the evolution of Robby’s strategies?
Try different rules for selecting the parents of the next generation. What leads to the fastest evolution? Is there ever a tradeoff between fast evolution at the beginning and how effective the winning strategies are at the end?
Try using spatial contiguity as a factor in selecting which individuals mate.
Try to train Robby to perform a somewhat more difficult task.
The run
command is key here. A chromosome is a list of strings, where each string is a procedure name. Calling run
on that string runs the procedure.
To represent strategies compactly, we use Unicode symbols (such as arrows).
To measure the similarity between different strategies, we use the highlevel list primitives map
and reduce
.
Simple Genetic Algorithm
Robby was invented by Melanie Mitchell and described in her book Complexity: A Guided Tour (Oxford University Press, 2009), pages 130142. Robby was inspired by The "Herbert" robot developed at the MIT Artificial Intelligence Lab in the 1980s.
This NetLogo version of Robby is based on Mitchell's earlier versions in NetLogo and C. It uses code from the Simple Genetic Algorithms model (Stonedahl & Wilensky, 2008) in the NetLogo Sample Models Library.
Robby resembles a simpler version of Richard E. Pattis' Karel the Robot, https://en.wikipedia.org/wiki/Karel_(programming_language).
If you mention this model or the NetLogo software in a publication, we ask that you include the citations below.
For the model itself:
Please cite the NetLogo software as:
Public Domain: To the extent possible under law, Uri Wilensky has waived all copyright and related or neighboring rights to this model.
(back to the NetLogo Models Library)