Fig 1.
Schematic of the block-catching task in the simulation.
A block is falling from top to bottom, either towards the left or the right. The animat has two sensory nodes, which are activated when the block is above them. It also has two motor nodes, which allow it to either move to the right or left. Depending on the size of the block, the animat’s task is to either catch or avoid the block. Reproduced from [47].
Table 1.
Descriptive statistics of the data.
Note that the number of perfect animats in both tasks are substantially larger than in Albantakis et al. [47]. This is likely due to optimization of the MABE framework, used for the evolutionary simulation, between then and now.
Fig 2.
Distributions of Φ and surprisal at the last generation of evolution depicted with half violin plot and half boxplot.
Distributions are shown for each task and for the hard task animats that evolved perfect fitness. The gray dots with horizontal lines represent the calculated average of the distributions.
Fig 3.
Change in average fitness (A), Φ (B) and surprisal (C) over evolutionary time for the easy task (black), the hard task (blue) and perfect animats in the hard task (red). The variables are first averaged across all values within each animat, then averaged across all LODs at each generation. The ribbons around the lines show the standard error from averaging across LODs.
Fig 4.
Average Φ and surprisal over trial time for the easy task (black), the hard task (blue) and perfect animats in the hard task (red).
A) average surprisal over trial time. B) average surprisal over trial time, centered around the first observation of the block. C) average Φ over trial time. D) average Φ over trial time, centered around the first observation of the block. Shading around lines is the standard error of the mean. See S5 Fig for the variability across LODs. Note that the averages on B and D are based on a decreasing number of trials as the relative timestep gets further away from 0. In order for the relative timestep to go to 30, at a given trial, the animat needs to see the block within the first few real timesteps.
Fig 5.
A-F) Within-trial representation of example animat behavior, Φ and surprisal. The gray box of each plot shows information about the trial and the animat, and the correlation between Φ and surprisal. The x-axis shows timesteps within the trial. The orange line denotes surprisal, and the green line denotes Φ. Each is plotted on an arbitrary y-axis for better comparison where only the zero point is indicated. The lower half of each plot shows the right and left sensory states (SR, SL) and motor states (MR, ML), with black lines indicating activation. Sensory states are active when the block is perceived. Motor states are activated by the animat, and makes it move in the given direction; if both motor states are active the animat stands still. Plots are chosen to show a variety of patterns of behavior, Φ and surprisal. All examples are taken from animats in the hard task.
Fig 6.
The distribution of correlation strengths between Φ and surprisal, across all trials on the last generation of all LODs.
Shown for time lags around 0, denoted above each of the graphs. Correlations are shown for all animats in the easy task (black), for all animats in the hard task (blue) and for only perfect animats in the hard task (red). The lagging variable was Φ, meaning that for negative lags the correlations measure the relationship between Φ and future surprisal, and vice versa for positive lags. Note that trials where Φ = 0 throughout are excluded from this analysis.
Fig 7.
Change in fitness (A), Φ (B) and surprisal (C) over evolutionary time averaged across LODs sorted into groups depending on animat correlation profiles at last generation. The groups are: Negative (purple), neutral (grey) and positive (brown) correlations.
Fig 8.
Average fluctuations in Φ and surprisal on trial time split into trials with negative (purple, coef. < -0.1, N = 4489), neutral (grey, coef. in (-0.1, 0.1), N = 1532) and positive (brown, coef. > 0.1, N = 5611) correlations between Φ and surprisal. A) average surprisal over trial time. B) average surprisal over trial time, centered around the first observation of the block. C) average Φ over trial time. D) average Φ over trial time, centered around the first observation of the block. Shading around lines is the standard error of the mean.
Fig 9.
A summary of the procedure for calculating Φ.
Fig 10.
Empirical goal priors (EGPs) of the LODs in the hard task, which reached perfect fitness.
For each perfect animat, 4 EGPs were constructed based on distinguishing first, catch and avoid trials, but also, left and right trials. The top row of blocks represent the avoid EGPs and the bottom row the catch EGPs. Those based on left trials are on the left and those based on the right trials are on the right. Each block consists of 8 heat maps, one for each of the LODs. The heat maps have 33 columns, one for each timestep (minus the first and last), and 4 rows, one for each possible sensory state. The sensory states are denoted by two binary digits, the first representing the left sensor and the second the right sensor (0 = off, 1 = on). Each cell in the heatmap represents the probability of the given sensory state at the given timestep. The color coding is constructed such that white represents a 0.25 probability. As the probability decreases below 0.25 colors become increasingly gray, and as the probability increases above 0.25 colors become increasingly orange. Probabilities with a value of 1 have a dark purple color to indicate states of certainty.