Science Corner: Companion to Three Chromosomes in a Trench Coat

If you haven’t been following along with Julie Reynolds‘ delightful series on phenotypic plasticity, I recommend catching up on that first. Julie shared some great real world examples, but not everyone has the opportunity to study overwintering insects like she does. So I thought I’d give you a hands-on example, albeit a simulated one. I’ve introduced my Quandary Den before. Briefly, players have to ‘zap’ or ‘tag’ robots for points, but the players have to evolve their gameplay approach. The versions I’ve shared before do not have any capacity for phenotypic plasticity, but we can change that.

We can actually keep the genotype (a series of letters) and the phenotype (a series of moves in the game) the same and just change how we map the one to the other. Previously, we read the genotype sequence as a sequential series of moves – go left, go down, zap right, and so on. As a result, the phenotype will be the same regardless of where the robots are; if we put the robots in new spots but keep the players’ genotype the same, the players will still make the same moves and possibly not do very well. Some examples are shown below; even just small shifts in the robot positions renders a previously perfect player genotype woefully ineffective.

To introduce more flexibility, we can instead interpret the genotype as a strategy for various game scenarios. We’ll define the scenarios as what the player can ‘see’ in each direction: up, down, left and right. So if the player is in the lower left corner with a robot above them, the scenario is robot, wall, wall, nothing. We can list every possible scenario and put them in some order of our choosing. Then we can read the genome as indicating in order which action to take in each of those scenarios. Once the game starts, we figure out which scenario the player is starting in, look up the action for that scenario and perform it, then recalculate the scenario and repeat. This way, the same genotype can result in a variety of phenotypes depending on where the robots are. Below are examples; the genotype is the same and so the strategy is the same, but the specific actions at each moment depend on the scenario or context, allowing the player to be successful even when the robots are moved.

I’ve implemented this version for you to experiment with. You can choose between the original deterministic genotype/phenotype match and the new contextual one. You can also change the other variables we’ve looked at before, like whether neutral mutations can accumulate or whether there can be multiple players. You can also decide whether the robots start in the usual place or a random one, and whether their position changes after each generation or not.

For some initial exploration, I’ve run a few experiments and plotted the fitness results in the chart at the top of the post. Dashed lines are for the deterministic genotype/phenotype map (‘D’) and solid lines are for the contextual one (‘C’). Standard starting positions are dark blue (‘Std’); random but fixed between generation starting positions are purple (‘Ran’); random and varying between generation starting positions are in light blue, green and yellow lines, averaging over 16, 32, or 64 different positions each generation (‘Var16’, ‘Var32’, ‘Var64’). I averaged over multiple starting positions each generation when the positions were random and varying in order to smooth out any impact of a particularly good or bad random arrangement. And I was curious if the number of different positions would matter; the effect seems small, but having to solve more positions seems to be harder.

There are a few other things we can note from these results. First, it seems that my standard starting positions (‘Std’) present a particularly easy challenge compared to random fixed positions (‘Ran’). It also looks like the deterministic approach has the advantage when the starting positions don’t change, getting to higher fitness faster; I’ll elaborate on why I think that is in a bit. However, the situation flips when the positions change each generation; in that case, a contextual strategy can get better results more quickly. And therein lies at least part of an answer for why it is worthwhile for organisms to have some phenotypic flexibility: dealing with changing, unpredictable conditions. For the curious, you can also check out the average number of chromosomes that evolved in each experiment; feel free to share your observations in the comments.

In addition to illustrating phenotype plasticity, this change brings up some other aspects of evolutionary biology. First is differences in fitness landscapes. The evolutionary process tends to select for populations with higher average fitness, which can be visualized metaphorically like climbing a hill with the top corresponding to the highest possible fitness. That hill image prompts the further metaphor of a fitness landscape. The simplest fitness landscape has a single, tall hill surrounding be plains that slope away from the hill in all directions. But just as with actual terrain, other fitness landscapes are possible, ones with multiple hills of varying heights and steepness. Changing to the contextual genotype/phenotype map results in a more varied fitness landscape, with hills that have peaks that are not as high as other hills’ but that have valleys in between. Consequently, if the player evolved a genotype corresponding to the peak of one of those low hills, mutations with lower fitness would have to accumulate before a higher hill could be climbed. By contrast, with the deterministic genotype/phenotype map the evolutionary process can keep taking steps upward (or sidewise, for neutral mutations) and ultimately get to the highest possible peak. This may explain the advantage the deterministic version has in the fixed starting position experiments.

The other evolutionary topic raised by changing genotype/phenotype map is the consistency of selection. When we keep the robots in the same place throughout the evolutionary process, we give the players an opportunity to learn those specific locations. With the contextual genotype/phenotype map, it makes more sense to move the robots around to see if we can evolve more general strategies that apply to a wide range of positions. (Although in hindsight, there are solutions in the deterministic version that would apply to multiple robot positions as well.) In a biological context, this would correspond to changing environments and changing selection pressures. Take for example the celebrated peppered moths; after increases in air pollution from the Industrial Revolution, darker moths survived better because they blended in with now-darker tree trunks, but subsequent clean air initiatives reduced pollution and the lighter moths became the most common again. Or consider the recent story about elephant populations without tusks, having evolved under selection pressure from poaching; it is expected that if poaching is brought under control, tusked elephants will rebound.

Fluctuating environmental conditions can lead to long-term consistency in populations even if there are short-term changes. Thus lasting population change is more likely if there is consistent selection in a particular ‘direction.’ This can happen when new niches are created or open up, such as when the atmosphere became oxygen-rich or when dinosaurs went extinct. This may partly explain why the fossil record shows periods of little or no change and periods of rapid change. When the environment is on average the same over time, the organisms living in it will tend to be the same as well. But when there is a persistent change in the environment, there are opportunities for new organisms.