/2015_GECCO_Bongard

Contains all of the material for replicating the paper entitled 'Evolving robot morphology facilitates the evolution of neural modularity and evolvability'

Primary LanguageC++

2015_GECCO_Bongard

Contains all of the material for replicating the paper entitled 'Evolving robot morphology facilitates the evolution of neural modularity and evolvability'

The paper itself is available here: http://www.cs.uvm.edu/~jbongard/papers/2015_GECCO_Bongard.pdf

This readme file will help you to replicate the results from the paper. This is done in a gradual fashion: each step enables you to replicate one of the figures or tables in the paper.

Step 1: Replicate Figure 1.

  • Click on Fig1.pdf in this directory.

Step 2: Replicate Figure 2.

  • This figure shows an evolved robot from experiment set 3 in Table 1. This robot was produced by allowing both the neural network and morphology of the robot to evolve. A bi-objective fitness function is used that only selects for age and grasping (the AFPO optimization method was used here).

  • Enter the Evolve_For_Age_And_Grasping directory.

  • Compile the C++ code by running ./makeModularity

  • This produces an executable file called Modularity.

  • Run this code at the command line with the following three arguments:

  • Modularity 0 1 1

  • This tells the code to run using random seed 0 (the first argument);

  • to select for robot controllers that settle into a fixed attractor (the second argument); and

  • to evolve the robot's morphology along with its controller (the third argument).

  • As the file runs, it will print out the percent of progress that has been achieved so far. When progress reaches 100%, the program will finish.

  • If the code is taking too long on your computer, open constants.h and reduce MAX_GENERATIONS. Recompile and re-run.

  • When the code finishes, results are stored in the Data/ directory.

  • Open Data/results_0.txt to see the results. Each row reports the modularity (m) and grasping ability (g) of the best robot in the population up to that point. Disregard the other data for now.

  • Now let's visualize the behavior of the best robot found by the end of your run. To do so, copy the Robot_Matrix files found in Data/ into the Visualization/ directory.

  • We will use Python, NumPy, SciPy and MatplotLib to do the visualization. If you are missing any of these, install them now.

  • In Visualization/ type 'python Robot_Draw.py'. This will re-create something similar to Figure 2.

  • Your visualization may not contain all four panels. This indicates that, by the end of your run, no robot had succeeded in all four environments. If this is the case, go back to constants.h and extend the length of the run by increasing MAX_GENERATIONS, re-compiling, and re-running.

Step 3: Replicate Figure 3.

This figure shows an evolved robot from experiment set 7 in Table 1. This robot was produced by allowing both the neural network and morphology of the robot to evolve. A tri-objective fitness function is used that selects for age, grasping ability, and behavioral conservatism.

  • Enter the Evolve_For_Age_Grasping_Conservatism directory.

  • Compile the C++ code by running ./makeModularity

  • Run this code at the command line: ./Modularity 0 1 1

  • When the run finishes, visualize the behavior of the best robot by copying the Robot_Matrix files in the Data/ directory into the Visualization/ directory and typing 'python Robot_Draw.py' there.

Step 4: Replicate Table 1.

  • Table 1 outlines the remaining runs that need to be performed to generate the remaining figures in the paper.

  • To perform 100 runs of experiment 1, enter the Evolve_Simple_Controllers/ directory.

  • Compile the code by running ./makeModularity.

  • Perform the first run using random seed 0: ./Modularity 0.

  • Perform the second run using random seed 1: ./Modularity 1.

  • Continue until 100 runs have been performed, which should generate 100 results files in the Data/ directory: results_0.txt, results_1.txt, ... results_99.txt

  • To perform 100 runs of experiment 2, enter the Evolve_For_Age_And_Grasping directory.

  • './Modularity 0 1 0' will perform the first run. The zero supplied for the third argument keeps the morphology fixed.

  • Perform 99 more runs: './Modularity 1 1 0', ... './Modularity 99 1 0'.

  • To perform the 100 runs of experiment 3, run './Modularity 100 1 1', ... '199 1 1'. Using random seeds 100 through 199 makes sure that the results files do not overwrite those generated from the previous step.

  • To perform experiment sets 4 and 5, enter Evolve_For_Age_Grasping_Modularity and perform './Modularity 0 1 0', ... './Modularity 99 1 0', './Modularity 100 1 1', ..., './Modularity 199 1 1'.

  • To perform experiment sets 6 and 7, enter Evolve_For_Age_Grasping_Conservatism and perform './Modularity 0 1 0', ... './Modularity 99 1 0', './Modularity 100 1 1', ..., './Modularity 199 1 1'.

Step 4: Replicate Figure 4a.

  • Open file Evolve_For_Age_And_Grasping/Data/results_0.txt. This file reports the results from the first run of experiment set 2. The last line of this file reports information about the best robot in the population when this run terminated.

  • The entry 'aa: x' reports the number of 'a'ttractors 'a'ttained by this robot. In other words, it reports how many of the 60 environments this robot can succeed in.

  • Write a python script that collects these number from each of the 100 files results_0.txt, ..., results_99.txt.

  • Compute the mean, and standard error of the mean, of these 100 numbers.

  • Do the same thing for the 100 numbers stored in results_100.txt, ... results_199.txt. This gives you the mean, and standard errror of the mean, for the performance of the 100 best robots produced when their morphologies were evolved as well (experiment set 3).

  • Plot these two means and SEMs. This will give you Figure 4a.

Step 5. Replicate Figure 4b.

  • Do the same as in step 4, but in this case extract the numbers associated with 'm: x', which are the neural modularities of these robots. Plot the resulting means and SEMs.

Step 6. Replicate Figure 4c.

  • Do the same as in step 4, but in this case extract the numbers associated with 'as: x', which are the behavioral conservatisms of these robots. Plot the resulting means and SEMs.

Step 7. Replicate Figure 4d.

  • Do the same as in step 4, but in this case extract the numbers associated with 'nd: x', which are the pose differences of these robots. Plot the resulting means and SEMs.

Step 8. Replicate Figure 5.

  • Repeat steps 4 through 7, but using the Evolve_For_Age_Grasping_Modularity/ directory.

Step 9. Replicate Figure 6.

  • Repeat steps 4 through 7, but using the Evolve_For_Age_Grasping_Conservatism/ directory.

Step 10. Replicate Figure 7.

  • Replot the data from Figs 4a, 5a, and 6a.

Any questions should be directed to jbongard@uvm.edu.