4 Block Stalagmites

4 Block Stalagmites preview image

3 collaborators

Uri_dolphin3 Uri Wilensky (Author)
Josh3 Josh Unterman (Author)
Dor Abrahamson (Author)

Tags

probability 

Tagged by Reuven M. Lerner over 11 years ago

problab 

Tagged by Reuven M. Lerner over 11 years ago

Model group CCL | Visible to everyone | Changeable by group members (CCL)
Model was written in NetLogo 5.0.4 • Viewed 712 times • Downloaded 86 times • Run 2 times
Download the '4 Block Stalagmites' modelDownload this modelEmbed this model

Do you have questions or comments about this model? Ask them here! (You'll first need to log in.)


WHAT IS IT?

4-Block Stalagmites is part of the ProbLab middle-school curricular material for learning probability. Related materials are a random generator called a "marbles scooper" and a sample space called a "combinations tower". In classroom activities, students working with 4-Block Stalagmites will have interacted with the marbles box and built its sample space and thus would have inferred expectations as to outcome distributions in hypothetical experiments with the marbles box.

4-Block Stalagmites is designed to enable users to experience insights into the binomial phenomenon and in particular to witness and understand the emergence of experimental outcome distributions that, by and large, are both consistent and in accord with proportions in the sample space. The model includes an interactive simulation of a binomial experiment with a sample size of four, which is comparable to an experiment of tossing four coins over and over, only that the four "coins" can land on green or blue, not heads or tails, and each "coin" has a fixed position in a 2-by-2 table that we call a 4-Block, unlike a set of four coins that has not inherent structure and lands "all over the place" on the table.

A unique feature of the model is that the outcome distribution is composed of the actual experimental samples themselves that are stacked one above the other in their corresponding columns. This is different from classical histograms that do not record which specific samples were taken but only their aggregate properties. For example, the particular 4-Blocks sampled appear in the distribution, rather than as just a record of the fact that a 4-Block with 3 green and 1 blue squares (in any order) were sampled.

The outcome distribution is in the form of "stalagmites" of stacked samples that have "dripped" down into their correct column. This creates a picto-graph histogram that grows bottom-up like a stalagmite. When the probability in the model is set at 0.5, this stalagmite will grow to 1:4:6:4:1 proportions. For other p values, the stalagmite will be tailed.

PEDAGOGICAL NOTE

There are four unique 4-blocks that each has exactly one green square, but there are six unique 4-blocks with exactly two green squares each. So, for a p value of .5 (when independent squares are equally likely to be green or blue), it is 1.5 times more likely to draw a two-green 4-block than a one-green 4-block (the ratio value of 6 to 4 is 1.5). This is worthy of attention, because students often need help in understanding how permutations are relevant to combinatorial analysis and, moreover, how combinatorial analysis is relevant to predicting the outcome distribution. When the sorting and coloring effects are activated as the simulation is running, the visual effect of the growing stalagmite is as though it is "stretching" the sample space of 16 elemental events into an outcome distribution of about 160 samples. Within columns we always expect all the elemental events to occur as frequently. However, this is true between columns only for a p value of .5. Note that we can change the p value and thus affect the overall shape of the stalagmites. For example, for the p value of .6, the two-green 4 Blocks will not occur as often as they would for a p value of .5, but the three-green 4 Blocks will occur more often than for a p value of .5. Due to the specifics of this change, the two-green and three-green columns are anticipated to be equally tall.

HOW IT WORKS

4-Blocks are randomly generated by asking each square to choose a color with a 'probability-to-be-target-color' chance of being green. Each 4-Block sample "drips" down one of the five columns in accord to its number of green squares. The stalagmite distribution can be sorted by type, even as it grows. There are 16 unique outcomes, so sorting the experimental outcomes by type results in 16 groups. These 16 groups are typically of uneven size, even for the p value of .5, but most often their sizes revolve around the average. For example for 160 samples taken, most groups will contain roughly 10 outcomes. You can "paint" these groups to enhance their visual groupiness.

HOW TO USE IT

Buttons: SETUP-initializes the View, essentially "emptying" the columns, and resets the variables and monitors. GO ONCE-generates a single 4-Block and sends it down its respective chute, whereas GO does so forever until one of the columns reaches the top of the display. GO-ORG-begins a run in which the samples sort themselves by type (see SORT OUTCOMES, below) ORGANIZE -rearranges outcomes within each column so that identical 4-blocks are grouped; DISORGANIZE -undoes this rearrangement. PAINT-colors outcomes by type so that identical 4-blocks appear of uniform color (the colors themselves are arbitrary-there is no inherent meaning or scaling); UNPAINT-returns the 4-blocks to their original appearance.

Switches: KEEP-REPEATS?-when Off, repeated outcomes are discarded from the Stalagmite. For example, say the simulation has already generated a 4-block with a single green square in the top-left corner. Any time later in the run, if the simulation generates another identical 4-block, it will descend the column and then disappear the moment it hits the stalagmite. But a 4-block with a single green square in the bottom-left corner would be kept, if it had not been generated. When On, repetitions are kept (as in standard outcome distributions). STOP-AT-ALL-FOUND?-when On, the run will end as soon as all 16 unique outcomes of the sample space have been randomly sampled. When Off, the run will continue until one of the columns reaches the top of the display. MAGNIFY?-when 'On,' a blown-up version of newly created 4-blocks is displayed to the side of the column. This helps, because the samples themselves are small and move fast. When Off, no blown-up sample is displayed.

Slider: PROBABILITY-TO-BE-TARGET-COLOR-determines the chance that each independent square in a 4-block will be green. For example, a value of 50 (50% or .5) means that each square has an equal chance of being green or blue, whereas a value of 80 means that each square has a 80% chance of being green and 20% chance of being blue. Monitors: EVENTS FOUND-keeps track of how many of the 16 possible 4-block outcomes have been randomly sampled.

Plot: EVENTS BY NUMBER OF OUTCOMES-shows how the sixteen elemental events are distributed by the number of outcomes sampled for each. When the first sample is taken, that event would be a '1' whereas all the other fifteen events are still at zero.

THINGS TO NOTICE

Setup the model in its default settings (with the 'probability' slider set to the value of 0.5 and the 'magnify?' switch set 'On'), slow down the model, using the speed slider above the View, and press 'Go'. See how a random 4-block sample is generated at the top of the View, just to the left of the stalagmite columns. Count up the number of green squares in this 4-block and see that the 4-block descends down a column bearing the corresponding numeral at the bottom. For example, if there are exactly two green squares in the random 4-block, it will go down the column with a "2" at the base.

Keep running the model slowly. See how samples are stacked on top of each other in the columns. Look closely at these samples and see if you can locate repeated outcomes, for example see if the 4-block with exactly two green squares in a particular diagonal formation occurred at least twice.

THINGS TO TRY

Set KEEP-REPEATS? to 'On' and STOP-AT-ALL-FOUND? to 'Off'. Press 'Go.' The columns will fill up until one of them hits the top, causing the run to stop. Compare the heights of the columns. What might you say about the relationship between these heights? Repeat this experiment and see whether any general pattern recurs.

Press SETUP then GO and wait until the run ends. Now press ORGANIZE. What happened? Press DISORGANIZE and then ORGANIZE, and watch the effect on the outcomes in the columns. Now press PAINT under each of the ORGANIZE and ORGANIZE conditions. When the outcomes are both organized and painted, what can you say about the relation among the sizes of the colored groups? That is, over repeated trials, is there any pattern in the relative sizes of these groups, or is it completely arbitrary?

Set KEEP-REPEATS? to 'Off' and STOP-AT-ALL-FOUND? to On. When you press GO the model will keep running until it has randomly sampled all of the unique outcomes in the sample space. How many samples, on average, are required in order to fill the entire sample space? Does this number change according to the settings of the probability? For example, if the probability is set at 80%, does it take as many trials to fill the sample space as compared to a setting of 50%? If not, why not? How about the extreme cases of 0% or 100%?

EXTENDING THE MODEL

Add monitors and/or graphs to explore aspects of the experiments that are difficult to see in the current version. For instance: o How many trials does it take for the experiment to produce an all-green 4-block? How is this dependent on the various settings?

o Are there more samples with an even number of green squares as compared to those with an odd number of green squares?

o How symmetrical is the set of stalagmites? How would you define "symmetry?" How would you quantify and display its changes over time?

NETLOGO FEATURES

RELATED MODELS

Some of the other ProbLab (curricular) models, including SAMPLER-a HubNet Participatory Simulation-feature related visuals and activities. In Stochastic Patchwork and especially in Sample Stalagmite you will see larger blocks, such as an arrays of green and blue squares. In the Stochastic Patchwork model and especially in 9-Blocks model, we see frequency distribution histograms. These histograms compare in interesting ways with the shape of the stalagmites in this model.

CREDITS AND REFERENCES

Thanks to Dor Abrahamson for the design and of this model as well as the implementation of the original model. Thanks to Josh Unterman for implementing the advanced procedures. This model is a part of the ProbLab Curriculum, originally under development at Northwestern's Center for Connected Learning and Computer-Based Modeling and now also at the Embodied Design Research Laboratory at UC Berkeley. For more information about ProbLab, please refer to http://ccl.northwestern.edu/curriculum/ProbLab/.

HOW TO CITE

If you mention this model in a publication, we ask that you include these citations for the model itself and for the NetLogo software:

  • Abrahamson, D. and Wilensky, U. (2006). NetLogo 4 Block Stalagmites model. http://ccl.northwestern.edu/netlogo/models/4BlockStalagmites. Center for Connected Learning and Computer-Based Modeling, Northwestern Institute on Complex Systems, Northwestern University, Evanston, IL.
  • Wilensky, U. (1999). NetLogo. http://ccl.northwestern.edu/netlogo/. Center for Connected Learning and Computer-Based Modeling, Northwestern Institute on Complex Systems, Northwestern University, Evanston, IL.

COPYRIGHT AND LICENSE

Copyright 2006 Uri Wilensky.

CC BY-NC-SA 3.0

This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-sa/3.0/ or send a letter to Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.

Commercial licenses are also available. To inquire about commercial licenses, please contact Uri Wilensky at uri@northwestern.edu.

Comments and Questions

Click to Run Model

globals [
         ;; colors of the background of the view and of the two possible colors in samples
         background-color column-color target-color other-color
         num-columns ;; how many columns there are in the graphics-window histogram
         num-target-color  ;; how many of the squares (patches) in the sample are of the favored color
         sample-right-xcor ;; the x-coordinate of the moving sample (not the magnified sample)
         sample-location-patch-agentset ;; patches where the moving sample will sprout
         token-sample-dude ;; bottom-left turtle in the moving sample
         stop-all? ;; Boolean variable for stopping the experiment
         side ;; tells how big the side of the block is--2 for 4-block

         popping?
         ]

patches-own [ column ]

breed [ column-kids column-kid ]
column-kids-own [ binomial-coefficient sample-list ]

breed [ sample-dudes sample-dude ]
sample-dudes-own [ distance-for-jump ]

breed [ baby-dudes baby-dude ]

;; jwu - instead of having the sample-dudes stamp, they're going to create
;; a sample-organizer. the sample-organizers are going to have a better idea
;; of which specific sample the sample-dudes represented.
breed [ sample-organizers ]
sample-organizers-own [
  sample-values
  original-pycor
]

to Go-org
  if stop-all? [stop]
    super-go
    organize-results
end 

to super-go
  if stop-all? [stop]
  ifelse popping? [
    no-display
    unpop
    go
    pop
    display
  ] [
    go
    display
  ]
end 

;; the popping? global controls the popping visuals

to pop
  set popping? true
  recolor-columns
end 

to unpop
  set popping? false
  recolor-columns
end 

;; jwu - different color for each sample-summary-value

to-report popping-color ; sample-organizers procedure
  report 15 + ((sample-summary-value * 10) mod 120)
end 

to-report sample-summary-value ; sample-organizers reporter
  let result 0
  let power-of-two 3
  foreach sample-values [
    if ? = 1 [
      set result result + 2 ^ power-of-two
    ]
    set power-of-two power-of-two - 1
  ]
  report result
end 

to-report sample-patches ; sample-organizers procedure
  let result []
  foreach n-values side [?][
    let i ?
    foreach n-values side [?] [
      set result lput patch-at ? (- i) result
    ]
  ]
  report result
end 

to display-sample ; sample-organizers procedure
  let patch-popping-color popping-color
  (foreach sample-values sample-patches [
    ask ?2 [
      ifelse popping? [
        set pcolor patch-popping-color
      ] [
        ifelse ?1 = 1 [
          set pcolor target-color
        ] [
          set pcolor other-color
        ]
      ]
    ]
  ])
end 

to recolor-columns
  reset-column-colors
  ask sample-organizers [
    display-sample
  ]
end 

to reset-column-colors ; column-kids procedure
  ask column-kids [
    ask patches with [ pcolor != black and
                       (pxcor = [pxcor] of myself or
                        pxcor = [pxcor] of myself - 1) ] [
      set pcolor [pcolor] of myself
    ]
  ]
end 

to make-a-sample-organizer ; sample-dudes procedure
  hatch-sample-organizers 1 [
    ht
    set sample-values map [ ifelse-value ([color] of ? = target-color) [1] [0] ]
                          sorted-sample-dudes
    display-sample
  ]
end 

to organize-results
  ask sample-organizers [
    if original-pycor = 0 [
      set original-pycor pycor
    ]
  ]
  ask column-kids [
    organize-column
  ]
  recolor-columns
end 

to organize-column ; column-kids procedure
  let column-organizers sample-organizers with [ pxcor + 1 = [pxcor] of myself ]
  (foreach sort-by [ [ sample-summary-value ] of ?1 <=
                     [ sample-summary-value ] of ?2
                   ] [self] of column-organizers
           sort [ pycor ] of column-organizers
           [ ask ?1 [set ycor ?2] ])
end 

to disorganize-results
 ask sample-organizers [
   set ycor original-pycor
 ]
 recolor-columns
end 

to startup
  ;set total-samples
end 

;; This procedure colors the view, divides patches into columns of equal length ( plus a single partition column),
;; and numbers these columns, beginning from the left, 0, 1, 2, 3, etc.

to setup
  clear-all
  set-default-shape turtles "square big"
  clear-output
  set background-color white - 1;black
  set column-color grey
  set target-color green
  set other-color blue
  set side 2

  set popping? false

  ;; num-columns is how many columns (bars) there are in the graphics-window histogram.
  ;; We need side ^ 2 + 1 columns in a histogram. For example, 3-by-3 samples (9 patches)
  ;; have 10 -- that is, 3 ^ 2 + 1 -- different possible counts of target-color (0, 1, 2, 3, ...9).
  set num-columns ( side ^ 2 + 1)

  ;; determines the location of the sample array beginning one column to the left of the histogram
  set sample-right-xcor -1 * round ( ( num-columns / 2 ) * ( side + 1 ) )

  ;; assigns each patch with a column number. Each column is as wide as the value set in the 'side' slider
  ask patches
  [
    set pcolor background-color
    ;; The following both centers the columns and assigns a column number to each patch
    ;; We use "side + 1" and not just "side" so as to create an empty column between samples
    set column floor ( ( pxcor + ( ( num-columns * ( side + 1 ) ) / 2 ) ) / ( side + 1 ) )
    if column < 0 or column >= num-columns
      [ set column -100 ]
  ]

  ;; leave one-patch strips between the columns empty
  ask patches with
  [ [column] of patch-at -1 0 != column ]
  [
    set column -100  ;; so that they do not take part in commands that report relevant column numbers
  ]

   ;; colors the columns with two shades of some color, alternately
   ask patches
   [
     if column != -100
     [
       ifelse int ( column / 2 ) = column / 2
       [ set pcolor column-color ][ set pcolor column-color - 1 ]
     ]
   ]

  ;; This draws the  base-line and creates a sample-kids turtle at the base of each column
  ask patches with
  [ ( pycor = -1 * max-pycor + side + 3 ) and  ;; The base line is several patches above the column labels.
    ( column != -100 ) ]
  [
    set pcolor black
    if [column] of patch-at -1 0 != column   ;; find the leftmost patch in the column...
    [
      ask patch (pxcor + side - 1)  ;; ...then move over to the right of the column
                ( -1 * max-pycor + 1 )
        [ set plabel column ]
      ask patch (pxcor + floor (side / 2))  ;; ...then move over to the middle of the column
                ( -1 * max-pycor + 1 )
      [
       sprout 1
        [
          hide-turtle
          set color pcolor
          set breed column-kids
          set sample-list []

          ;; each column-kid knows how many different combinations his column has
          set binomial-coefficient item column binomrow (num-columns - 1)
        ]
      ]
    ]
  ]
  set stop-all? false
  set num-target-color false
  reset-ticks
end 

to go
  if stop-all? [stop]
  ;; The model keeps track of which different combinations have been discovered. Each
  ;; column-kid reports whether or not its column has all the possible combinations. When bound? is true,
  ;; a report from ALL column-kids that their columns are full will stop the run.
  if stop-at-all-found? [if count column-kids with [length remove-duplicates sample-list = binomial-coefficient] = count column-kids
    [stop]]
  sample
  ifelse magnify? [ magnify-on-side ] [ ask baby-dudes [ die ] ]
  drop-in-bin

  tick
  if plot? [ histogram-blocks ]
end 

;; This procedure creates a square sample of dimensions side-times-side, e.g., 3-by-3,
;; located to the left of the columns. Each patch in this sample sprouts a turtle.
;; The color of the sample-dudes in this square are either target-color or other-color,
;; based on a random algorithm (see below)

to sample
  ;; creates a square agentset of as many sample-dudes as determined by the 'side' slider,
  ;; positioning these sample sample-dudes at the top of the screen and to the left of the histogram columns
  set sample-location-patch-agentset patches with
  [
    ( pxcor <= sample-right-xcor ) and
    ( pxcor > sample-right-xcor - side ) and
    ( pycor > ( max-pycor - side ) ) ]

  foreach sort sample-location-patch-agentset
  [
    ask ?
    [
      sprout 1
      [
        ht
        set breed sample-dudes
        setxy pxcor pycor

      ;; Each turtle in the sample area chooses randomly between the target-color and the other color.
      ;; The higher you have set the probability slider, the higher the chance the turtle will get the target color
      ifelse random 100 < probability-to-be-target-color
        [ set color target-color ]
        [ set color other-color ]
      st
      ]
    ]
  ]
  ;; num-target-color reports how many sample-dudes in the random sample are of the target color
  set num-target-color count sample-dudes with [ color = target-color ]
end 

;; procedure in which the sample turtles create an enlarged duplicate on the left side of the screen and this enlarged sample makes
;; a large duplicate of the sample. This helps users see the samples that may otherwise be too small to see comfortably.
;; Samples are small for side = 3 because we want the entire sample space to fit into the view.

to magnify-on-side
  ask baby-dudes [die] ;; clears the way for new magnified sample
  ask sample-dudes
  [
    hatch-a-big-baby
  ]
end 

to hatch-a-big-baby ;; sample-dudes procedure
  hatch 1
  [
    set breed baby-dudes
    set size 12 * ( 8 - side ) / side
    ;; This is tricky.  We want to center the new turtles vertically, and
    ;; put them to the right of all the columns, with a little space between.
    ;; The code is complicated, because it is supposed to work for other x-blocks, too
    setxy -1 * ( size * .35 ) + ( sample-right-xcor + ( .35 * size * ( xcor - sample-right-xcor ) ) )
          ( ( side - 1 ) * size / 2 ) + ( .35 * size * ( ycor + min-pycor ) )
    set size size * .33
  ]
end 

;; This procedure moves the random sample sideways to its column and then down above other previous samples
;; in that column.

to drop-in-bin
  find-your-column
  descend
end 

;; The random sample moves to the right until it is in its correct column, that is, until it is in the column
;; that collects samples which have exactly as many sample-dudes of the target color as this sample has.
;; The rationale is that the as long as the sample is not in its column, it keeps moving sideways.
;; So, if the sample has 9 sample-dudes (3-by-3) and is moving sideways, but 6 of them are not yet in their correct column,
;; the sample keeps moving. When all of the 9 sample-dudes are the sample's correct column, this procedure stops.

to find-your-column
  ask sample-dudes [ set heading 90 ]
  while
  [ count sample-dudes with [ column = num-target-color ] != side ^ 2  ]
  [
    ask sample-dudes
    [ fd 1 ]
  ]
end 

;; Moves the sample downwards along the column until it is either on the base line or
;; exactly over another sample in that column.

to descend
  let lowest-in-sample min [ pycor ] of sample-dudes
  ask sample-dudes
  [ set heading 180 ]

  ;; The lowest row in the square sample is in charge of checking whether or not the sample has arrived all the way down
  ;; In order to determine who this row is -- as the samples keeps moving down -- we find a turtle with the lowest y coordinate
  ;; checks whether the row directly below the sample's lowest row is available to keep moving down
  set token-sample-dude one-of sample-dudes with [ pycor = lowest-in-sample ]
  while
  [
    ( [ [pcolor] of patch-at 0 -2 ] of token-sample-dude ) != black  and
    ( [ [pcolor] of patch-at 0 -2 ] of token-sample-dude ) != target-color and
    ( [ [pcolor] of patch-at 0 -2 ] of token-sample-dude ) != other-color
  ]
  [
    ;; As in find-your-column, shift the sample one row down
    ask sample-dudes
    [ fd 1 ]

  ;; Instead of establishing again the lowest row in the sample, the y coordinate of the row
  ;; gets smaller by 1 because the sample is now one row lower than when it started this 'while' procedure
  set lowest-in-sample ( lowest-in-sample - 1 )
  ]

  ;; Once sample-dudes have reached as low down in the column as they can go (they are on top of either the base line
  ;; or a previous sample) they might color the patch with their own color before they "die."
  finish-off

  ;; If the column has been stacked up so far that it is near the top of the screen, the whole supra-procedure stops
  ;; and so the experiment ends
  if max-pycor - lowest-in-sample < ( side + 1 ) [ set stop-all? true ]
end 


;; we can't sort by who number, because who numbers get reused in weird ways, it seems.

to-report sorted-sample-dudes
;report sort sample-dudes
report sort-by [
         (([pxcor] of ?1 < [pxcor] of ?2) and ([pycor] of ?1 = [pycor] of ?2)) or
         (([pycor] of ?1 > [pycor] of ?2))
       ] sample-dudes
end 

to finish-off
  ;; creates local list of the colors of this specific sample, for instance the color combination of a 9-square,
  ;; beginning from its top-left corner and running to the right and then taking the next row and so on
  ;; might be "green green red green red green"
  ;; jwu - need to use map and sort instead of values-from cause of
  ;; the new randomized agentsets in 3.1pre2
  let sample-color-combination map [ [color] of ? ] sorted-sample-dudes

  ;; determines which turtle lives at the bottom of the column where the sample is
  let this-column-kid one-of column-kids with [ column = [ column ] of token-sample-dude ]

    ;; make the upper left sample-dude create a sample-organizer
    let the-sample-sample-dude max-one-of (sample-dudes with-min [ pxcor ]) [ pycor ]

  ;; accepts to list only new samples and makes a previously encountered sample if keep-duplicates? is on
  ifelse not member? sample-color-combination [sample-list] of this-column-kid
  [
    ask the-sample-sample-dude [
      make-a-sample-organizer
    ]
    ask sample-dudes
    [ die ]
  ]
  [
    ifelse keep-repeats? [
    ask the-sample-sample-dude [
      make-a-sample-organizer
    ]
    ask sample-dudes
    [ die ]
    ] [
      ask sample-dudes
        [  die ]
    ]
  ]
    ask this-column-kid
    [ set sample-list fput sample-color-combination sample-list ]
end 

;; procedure for calculating the row of coefficients
;; column-kids needs their coefficient so as to judge if their column has all the possible different combinations

to-report binomrow [n]
  if n = 0 [ report [1] ]
  let prevrow binomrow (n - 1)
  report (map [?1 + ?2] fput 0 prevrow
                        lput 0 prevrow)
end 

;;if the model has been run, report the number of patches
;;with the target-color -- otherwise, display "--"

to-report #-target-color
  ifelse ticks != 0
  [ report count patches with [ pcolor = target-color ] ]
  [ report "--" ]
end 

;;if has been run, report the number of patches
;;with the other-color -- otherwise, display "--"

to-report #-other-color
  ifelse ticks != 0
  [ report count patches with [ pcolor = other-color ] ]
  [ report "--" ]
end 

;; reports the proportion of the sample space that has been generated up to now

to-report %-full
  ifelse samples-found = 0
    [ report precision 0 0 ]
    [ report precision ( samples-found / ( 2 ^ ( side ^ 2 ) ) ) 3 ]
end 

to-report samples-found
  report sum [ length remove-duplicates sample-list ] of column-kids
end 

to-report total-samples-to-find
  report precision ( 2 ^ ( side ^ 2 ) ) 0
end 

to histogram-blocks
  let sample-value-summaries [ sample-summary-value ] of sample-organizers
  let possible-values n-values (2 ^ (side * side)) [?]
  let results []
  foreach possible-values [
    let i ?
    set results lput length filter [? = i] sample-value-summaries results
  ]

  set-current-plot "Events by Number of Outcomes"
  let max-results max results
  if mean results > 0 [ set-plot-x-range 0 (max-results + 1) ]
  set-current-plot-pen "Histogram"
  histogram results
  set-current-plot-pen "Mean"
  let mean-results mean results
  plot-pen-reset
  plotxy mean-results 0
  plotxy mean-results plot-y-max
end 


; Copyright 2006 Uri Wilensky.
; See Info tab for full copyright and license.

There are 10 versions of this model.

Uploaded by When Description Download
Uri Wilensky over 11 years ago Updated to NetLogo 5.0.4 Download this version
Uri Wilensky about 12 years ago Updated version tag Download this version
Uri Wilensky about 12 years ago Updated to version from NetLogo 5.0.3 distribution Download this version
Uri Wilensky almost 13 years ago Updated to NetLogo 5.0 Download this version
Uri Wilensky over 14 years ago Updated from NetLogo 4.1 Download this version
Uri Wilensky over 14 years ago Updated from NetLogo 4.1 Download this version
Uri Wilensky over 14 years ago Updated from NetLogo 4.1 Download this version
Uri Wilensky over 14 years ago Updated from NetLogo 4.1 Download this version
Uri Wilensky over 14 years ago Model from NetLogo distribution Download this version
Uri Wilensky over 14 years ago 4 Block Stalagmites Download this version

Attached files

File Type Description Last updated
4 Block Stalagmites.png preview Preview for '4 Block Stalagmites' over 11 years ago, by Uri Wilensky Download

This model does not have any ancestors.

This model does not have any descendants.