Lifespan Distribution on a Torus

After simulating 175000 random patterns on a 20×20 torus, the distribution of the lifespans was as follows:

The bins in the graph have a width of 10, and the peak occurs in the lifespan 91 – 100 bin. Some other interesting stats include:

• Maximum observed lifespan: 2092
• Minimum observed lifespan: 14
• Average lifespan: 210
• Median lifespan: 164
• Standard deviation of lifespan: 159.3

In order to try to estimate the longest lifespan of any pattern on a 20×20 torus, I will assume that if the lifespan is above 100, then it follows an exponential distribution, which appears to be a reasonable assumption given the above histogram. Indeed, this leads to the following histogram, which has the exponential of best fit overlaid in black. The equation of the exponential is y = 9088.7(0.938)^x, and the R² value is 0.997 (that’s one heck of a good fit).

Now that we have our exponential of best fit, we can get an estimate of the maximal lifespan in this scenario by assuming that the lifespan of the 2^20*20 different patterns are independent of each other (this may not be a particularly good assumption, as two patterns, for example, will have the same lifespan if they are translations of each other — nonetheless it seems that the number of reductions that can be made by symmetry arguments like that is tiny compared to the total number of such patterns, so let’s roll with it). We then set

and solve for x to get a maximal lifespan of x = 4474. This seems believable and not particularly shocking to me given the statistics that I listed earlier.

Lifespan Distribution on a Plane

Typically, Life enthusiasts are more interested in results on the infinite plane than on a torus, as it leads to behaviour that is much more chaotic and unpredictable. For example, even though we just saw that the longest-lived 20×20 pattern on a torus likely lives for less than 4500 generations, there is a known 9×15 pattern that lives on the plane for over 29000 generations. We will now proceed analogously to the torus discussion above to see how much better we can reasonably expect to do. We begin with the distribution of the lifespan based on 60000 initial 20×20 patterns:

The bins in the graph have a width of 50, and the peak occurs in the lifespan 151 – 200 bin. The other basic stats are listed below, and as we would expect, the average lifespan and the standard deviation of lifespans are both much greater on the plane than they are on the torus.

• Maximum observed lifespan: 13612
• Minimum observed lifespan: 15
• Average lifespan: 678
• Median lifespan: 368
• Standard deviation of lifespan: 885.6

The main difference with the analysis in this situation is that instead of assuming the lifespan itself follows an exponential distribution, I will assume that some power of the lifespan follows an exponential distribution; the reason for this is that the R² is quite low if we try to fit an exponential to the distribution of the lifespans itself. Fitting an exponential to lifespan^0.75 seems to give the maximum R² value (0.9857), so this is what I will use. The following histogram shows the curve of best fit, which has the equation y = 1854.6(0.957)^x, where this time x = lifespan^0.75.

Finally, we follow our method from before and set

and solve for x to get a maximal lifespan of x = 120795. Of course, this number should be taken with a grain of salt, because that pattern I mentioned earlier that has a lifespan of over 29000 generations was found during a simulation of a whopping 12 billion 20×20 patterns, while our equation of best fit tells us that after 12 billion patterns, we would expect only to see a pattern with lifespan 6206 (when really it should only take about 400 patterns to see such a lifespan). Similarly, in the longest-lived patterns found by the Online Life-Like CA Soup Search, a pattern has been found with lifespan 24389. According to our formula, we should have to look at roughly 9 million billion billion billion soups before finding such a pattern, but only 40 million soups have actually been searched.

This all seems to suggest that the longest-lived pattern in a 20×20 box may in fact live considerably longer than 120000 generations, which goes to show just how in the dark we are when it comes to some of these simple questions — since the search space is so unimaginably huge, we likely will never know for sure what the true upper limit is.

Downloads:

• Lifespan raw data and graphs [.zip of an Excel spreadsheet — 2.91MB]
• Lifespan recorder scripts [.zip of two Golly Python scripts]

The solution to the given problem is to emulate the Game of Life not on an infinite planar grid, but rather on a toroidal grid; that is, we fix our 20×20 universe and treat the top row as if it borders the bottom row and the left row as if it borders the right row. Running the simulation in this set-up results in data that matches our intuition more closely:

The results are based on 13592 random patterns of each density 1%, 2%, 3%, …, 99%. As with the previous graph, we notice that the lifespan plateaus in the 25% – 50% range, providing evidence that our intuition for a maximum at 37.5% is probably well-founded. Also, at the request of Elithrion in the comments on the previous post, I have included the median plot for the lifespans this time as well, though the median seems to behave pretty similarly to the average (it’s just more or less uniformly 25% smaller). As we would expect, the average lifespan in the toroidal universe set-up is much lower than the average lifespan in the infinite planar grid set-up, because expanding patterns have nowhere to go. The longest-lived pattern that was found on the torus had a lifespan of 2169 (and an initial density of 30%), whereas the longest-lived pattern that was found on the infinite plane (over significantly fewer trials, I might add) was 14042.

While I’m presenting results, how about we take a look at the final population of patterns of various starting densities?

The behaviour is almost identical to that of the average/median lifespan. The one cool thing is that, in addition to the plateau centred at ~37.5%, here we see that the median final population peaks with a value of 13 at an initial density of 38%. Nevermind the fact that this is almost certainly a statistical anomaly — it helps support my original hypothesis, and if the popular news media has taught me anything, it’s that randomly picking out things like that and claiming that they are significant is always mathematically sound.

Downloads:

• Lifespan and final population raw data [.zip of an Excel spreadsheet — 2.99MB]
• Torus lifespan script [.zip of a Golly Python script]

As I mentioned in this post, one way to learn about the general behaviour of Conway’s Game of Life (or any other Life-like cellular automaton) is to generate random patterns and see what stable patterns they typically evolve into (if any). If you’ve played around with Conway’s Game of Life, then you more than likely are very familiar with blocks, blinkers, beehives, loaves, and maybe even pulsars, simply as a result of drawing random gibberish on the grid and letting it evolve. This is exactly the idea behind the Online Life-Like CA Soup Search, just on a larger scale — what if we had lots of people’s computers drawing random gibberish on Life grids, letting them evolve, and storing the resulting patterns in a communal database?

This method of collecting patterns in Life is known as a census, and to my knowledge it has been carried out on a relatively large scale twice in the past; once by Achim Flammenkamp (results here), and once by Andrej Okrasinski (results linked in the first paragraph here). As we would expect, the block is by far the most common still life, and the blinker is by far the most common oscillator. If we scroll down a bit, we see that the pulsar is the fourth most common oscillator (which is reasonably surprising, given its size). Scrolling much past that, we get into the interesting stuff — objects that no reasonable person has ever seen from drawing random stuff and seeing what it evolves into, simply because they occur so infrequently.

So why bother doing another census if two rather large-scale censuses (censi?) have already been carried out? My reasons are threefold:

1. There has never (to my knowledge) been a census of any other Life-like cellular automata conducted. The Online Life-Like CA Soup Search is completely general in that it can conduct a census of any Life-like rule (assuming that rule is stable, of course). Censuses have already been started for five other rules, including 2×2, Day & Night, HighLife, and Move. Because some of these other rules have been studied relatively little, this is an easy way of discovering new yet fundamental patterns in them. For example, it didn’t take long for the soup search script to uncover a c/8 line puffer in the 2×2 (B36/S125) rule; the very first known pattern that exhibits infinite growth in 2×2!
   

   Line puffer in 2x2

2. Previous soup searches were limited in that they only looked for very specific types of objects — usually still lifes and oscillators (though Achim Flammenkamp did a search for spaceships and puffers that can be viewed here). The Online Life-Like CA Soup Search searches for still lifes, oscillators, spaceships and puffers all at once, and even gets pseudo objects (such as the bi-block) as well.
3. Multiple users can run the script and contribute to the results simultaneously. Rather than just having one person run a script that they themself wrote, the Online Life-Like CA Soup Search works by having users download a script and that script uploads results to the central database. This obviously provides a huge speed-up, and allows the census to continuously improve and expand.

So what are you waiting for? Visit the Online Life-Like CA Soup Search, download the script, and put your computer’s CPU cycles to good use!

Census results in Conway's Game of Life

The question of how long soup generally lasts before stabilizing is quite vague, because there are two important considerations that we have not specified:

1. How dense should the soup be? Clearly soup with ON density of 5% will survive only a fraction as long as soup with ON density of 50%.
2. How large of a region of soup should we let evolve? Smaller regions will stabilize much more quickly than large regions, and infinite regions will likely never stabilize(?), not to mention we can’t effectively simulate an infinite field of random cells.

For #2, I will choose the region to be of size 20×20. This decision is pretty arbitrary, but part of the reason I chose such a small size is as a desire to make this script kill two birds with one stone; any particularly long-lived patterns of this size that are discovered can likely be considered methuselahs, while larger patterns can not.

For #1, we might naively expect that patterns will live longest if they have density 37.5%, since cells are born if and only if they have exactly three ON neighbours (out of a possible eight = 37.5%). To test this theory, I created a Golly script that creates regions of varying density and checks how long it takes for them to stabilize. The following graph shows the average lifespan of 20×20 patterns with densities ranging from 1% to 99%, based on 5000 generated patterns for each percentage point.

Indeed, it looks like the longest-lived density estimate of 37.5% isn’t too far off. The true maximum in this set-up might actually be slightly higher, but that is most likely caused by the same thing that is causing the hump centered at 90%: edge effects. Roughly speaking, because we are simulating a finite region of soup on an infinite grid (as opposed to on a toroidal universe), patterns of higher density will tend to expand out very quickly in their first few generations, giving them longer lifespans in general than they would have on a toroidal universe. Alas, Golly doesn’t support toroidal universes, so the toroidal lifespan results will have to wait for another day.

Update [July 8, 2009]: I have attached to the end of this post the Python script used to generate these results.

Update [July 11, 2009]: See this post, which deals with the “more ambitious project” that I mentioned.

Update [July 19, 2009]: See this post, which deals with soup longevity on a torus.

Downloads:

• Methuselah 14042 – A methuselah with lifespan 14042 found during the search [.zip of an RLE file]
• Methuselah Search Script [.zip of a Golly Python script]
• Soup Density Data [Excel spreadsheet — 26.0kB]