"Blocks" is a Java program that was written as an experiment in simulating natural selection. The replicators are little square shapes (the "blocks") and the environment they exist in is a rectangular grid on the computer screen.
The blocks are capable of four kinds of actions: moving, fighting, eating, and dividing. Each block has internal logic called a "decision procedure" that determines what action it will take, based on the conditions that it finds itself in.
When a block divides, the child block inherits the parent's decision procedure. However random mutations happen, so that the child doesn't always get an exact copy of the decision procedure of the parent. The hope is that, as the simulation runs, natural selection will operate - resulting in blocks that have more effective decision procedures.
Running the Simulation
The software comes as a Java "jar" file, called: blocks.jar
In order to run the program, you need to have the Java runtime environment installed on your computer. Once you have Java installed, you should be able to run the simulation by typing the command: java -jar blocks.jar
If you have any problem downloding or running the program please let me know.
The Components of the Simulation
Below is a screen shot of a simulation in progress. A number of the system components are visible: the Blocks Console, the Status Window, the Debug Window, and a Chart which graphs information about the characteristics of the block decision procedures.
All in all there are seven major components of the blocks system:
The Console and the World
When the simulation first starts, you will be confronted with a blank console screen and a debug window, which will look something like this:
The Console window contains menus at the top, and has a big rectangular area, which represents the "world" that the blocks live in. The first thing you are probably going to want to do is to go to the "Console" menu (on the left) and select the "Type 1 World" menu item. This will create a world of blocks and draw them on the console:
You will also probably see a Status Window pop up. That will be described later.
After that, go to the "World Operations" menu and select start. This will start the simulation running, and you should see the blocks moving around the screen, eating each other and so on. When blocks turn black it means that they have died, but not been eaten yet.
It's possible that the world you have just created might run for quite a while, or the blocks might all die out after only a short time. It you want to stop the world, go to the "World Operations" and select "Stop"
To exit the entire program, go to the "Console" menu and select "Exit Program"
Blocks and Decision Procedures
The way the simulation operates is that each block in turn gets a chance to "make a move" over and over again, as long as there are any live blocks left in the world.
When the block gets its move, it "looks around" to see what kind of other blocks it is touching or close to, and then it invokes its "decision procedure" to tell it what to do. For example, it might decide to try to "chase" another block, or "run away" from it, perhaps based on the size of the other block or whether or not they are the same color.
The simulation currently has four different kinds of decision procedures. When you select "Type 1 World" from the "Console" menu, you are creating a world of blocks that has the first type of decision procedure (referred to as DP1.) When you select the "Type 4 World" you get a DP4 decision procedure. DP1 is the simplest decision procedure and DP4 is the most complex.
The decision procedures are overall "frameworks" for decision making, but the exact decisions that get made depend on the specific "characteristics" of the individual block. These "characteristics" or "attributes" are randomly generated when the world is created. All blocks in a given world get the same decision procedure, but each color of block gets different "characteristics." When blocks divide, their offspring inherit their characteristics (but mutations occur at random, so that new characteristics are periodically introduced.)
DP1 and 2
DP1 and DP2 are very similar. When using one of these decision procedures, each block has the following four characteristics, which can take on the values indicated:
Characteristic Possible Values -------------- --------------- HOW_MOVES RANDOM, TOWARD_ALL, AWAY_ALL, TOWARD_SAME, TOWARD_OTHER, ... HOW_EATS NONE, OTHER, ALL HOW_DIVIDES NEVER, SOON, LATER DEFENDS_SELF TRUE, FALSESo, for example, a block that has HOW_MOVES = TOWARD_ALL will always "chase" all other blocks. One that has HOW_MOVES = RANDOM will simply wander around the screen. Some values of characteristics are obviously losers, from the natural selection perspective, such as HOW_DIVIDES = NEVER. We expect to see these values "die out" as the simulation progresses.
If you do a "left click" with the mouse on a block, various information about that block will be dumped to the debug window, including all the values of its decision procedure characteristics:
World 0: Block 10: Block 10: dpType=1, location=(152,83), energy=784, size=28, age=0 World 0: Block 10: dead=false, deleted=false World 0: Block 10: HOW_MOVES = TOWARD_OTHER World 0: Block 10: HOW_EATS = OTHER World 0: Block 10: HOW_DIVIDES = LATER World 0: Block 10: DEFENDS_SELF = TRUE World 0: Block 10: color = java.awt.Color[r=255,g=0,b=0] World 0: Block 10: closest block to me is 8 (distance is 3)
DP3
DP3 has a more sophisticated mechanism. It goes through five "stages" of decision making:
In order to make each of these decisions (other than the last one) there is a "weighted network", for which the inputs are the things that the block knows (the sizes, colors, and so on of the blocks involved.) When we are trying to select which block is interesting, the outputs are the candidate blocks themselves. When we are trying to decide which actions to do, the outputs are the potential actions.
In DP3, what is inherited by a child block is the "weights" from the decision networks. Here is part of the output of a left click on a type 3 block:
World 0: Block 13: Block 13: dpType=3, location=(155,176), energy=841, size=29, age=0
World 0: Block 13: dead=false, deleted=false
World 0: Block 13: Weights:
Stage[0].DIFFERENT = 10
Stage[0].BIGGER = 5
Stage[0].HUNGRY = 0
Stage[0].RELATED = 0
Stage[0].DEAD = 5
Stage[0].THRESHOLD = 7
Stage[1].DF_M = 1
Stage[1].BIG_M = 5
Stage[1].HGY_M = 10
Stage[1].REL_M = 5
Stage[1].DED_M = 4
Stage[1].DF_F = 10
Stage[1].BIG_F = 1
Stage[1].HGY_F = 1 ...
The output would go on for a while, until all the weights had dumped out.
Here is how to interpret the weights. For stage[0] (the first of the
five stages described above) each line tells how much weight the attribute
named contributes to the candidate block in question. So, it the block
being examined is a different color that the block making the decision,
then 10 will be added to that block's score. If it is bigger, then 5
will be added. When all the weights are added up, they have to exceed
the threshold (in this case 7) for that block to be a possibility as
the most "interesting" block. Whether it is actually selected or not
depends on whether its final score is higher than any of the other
blocks that were considered.
In stage 1 we have already selected a block and we are trying to figure out which of the 3 actions we are going to do (fight, eat, move.) The (rather cryptic) notations mean:
DF = Different BIG = Bigger HGY = Hungry (this means the block making the move is hungry) REL = The 2 blocks are "related" (one is the parent of the other) DED = The target block is dead M = move F = fight E = eatSo, in the output above, "Stage[1].DED_M = 4" means that in the second stage, the target block being dead adds a weight of 4 to the move action. "Stage[1].DF_F = 10" means that if the target block is a different color a weight of 10 is added to the fight action. In the end, the action with the highest score wins.
DP4
DP4 is similar to DP3 in that it has the same five stage decision process. In DP4 all 5 stages are determined by weighted networks. The networks in DP4 are a little bit more sophisticated than those of DP3. For example, the DP4 networks contain weights for both "true" and "false" values of the inputs. This means that, for example, one weight can be given to the condition that the target block is bigger and a separate weight can be given to the condition that the target block is not bigger. In the "left click" output, the "false" input weights are denoted by a tilde:
World 0: Block 11: Block 11: dpType=4, location=(231,69), energy=900, size=30, age=0
World 0: Block 11: dead=false, deleted=false
World 0: Block 11: stbNet:
SelectNet(ChooseTouching)
different---(-4)--> block
~different---(7)--> block
bigger---(7)--> block
~bigger---(5)--> block
related---(-8)--> block
~related---(7)--> block
dead---(-4)--> block
~dead---(2)--> block ...
It's easier to get a handle on what the decision networks are doing by
examining type 4 blocks. Although you can get a dump of the weights by
doing a left click, with a type 4 block there is a much better way.
You can do a Shift-Control-LeftClick and get a tool that lets you
visually examine and experiment with one of the block's decision networks.
The process goes like this:
The Status Window
Every time a world is created, an associated Status Window is displayed:
The Status Window displays various information about the running world. The most interesting counters are probably turns, moves, births, deaths, and mutations. A "move" happens every time each individual blocks gets to do something. A "turn" is defined as one "sweep" through all the blocks in the world, each block getting one move. A "birth" is counted every time a block divides, creating a new block. A "death" simply means that a block has died. This can happen either because it was killed by another block, or because it died of "old age."
Mutations tells you how many times the characteristics inherited by a child block have been randomly changed so that they are no longer exactly the same as the characteristics of the parent.
Debug Windows
Each object in the Blocks system has the ability to output debug messages to a specified Debug Window. By default one main debug window is created when the system starts up, and all messages go there. However, other windows can be created. For example, if you do a Shift-Left click on a block, it will create a separate Debug Window for that one block.
The Command Window
The things that can be accomplished with the menus on the Block Console are limited. To do more complicated things, you need to use the Command Window.
Go to the "Console" menu and select the "Command Window" item. When the command window comes up, go to the text field on the top and type "help" and you should see something like this:
Probably the most interesting command is "logWorld" which allows you to start up a world with a "logger" that keeps track of various information and presents it on graphical charts described later.
For example: logWorld 1 250 200 true foo.log
Will start up a "type 1" world with a logger. The logger will take a
snapshot of the state of the world (and graph it) every 250 turns. There
will a total of 200 snapshots on the graph. The "true" parameter means
that the logger will stop the world at the end of the 200 snapshots.
If we had put "false" here, then the logger would simply continue,
and the graph would become a "sliding window" showing the most recent
200 data points.
The last parameter is the name of a file to output log data to. You MUST give it the name of a file that DOES NOT exist yet, or the program will simply crash. This is a bug that needs to be fixed.
You will get different graphs depending on what type of world you are running.
Graphs
Whenever you have a logger running you will at least get this graph of "System Variables"
Keep an eye on the "Memory" chart. The current implementation appears to have problems with the JVM garbage collector, and you will likely see memory running out during a long run.
The other graphs will depend on what type of world you have. This is because the "inheritable attributes" of the different types of blocks (based on their decision procedures) need to be displayed differently.
This is an example of the DP1/2 Attributes graph you will see for either a type 1 or type 2 world:
Take a look at the DEFENDS_SELF attribute and notice that the number of blocks that have "FALSE" (they do not defend themselves) sinks down pretty quickly. No surprise here. Also, in HOW_DIVIDES, the number of blocks that never divide (the very small blue line in the lower left of that chart) dies out right away.
Once you get into the more sophisticated decision procedures it's not easy to come up with a "readable" graph of the average weights, so you might not be able to make too much of their graphs. I'm still looking for a better way to do this.
Final Note
This brief introduction does not touch on many aspects of the simulation. Please feel free to contact me if you would like further information.