Biology Simulation
The following videos show biologically motivated scenarios that demonstrate the evolution of predatory, mutualistic, and parasitic behaviors. Many of the observed behaviors appear surprisingly complex, which was unexpected given the extremely simple (6-neuron hidden layer) neural network that operates as the controller:
In the biological scenarios, rules were loosened to allow multiple robots to occupy the same space without colliding. The different robot types (red & green) were then given different maximum speeds and different fitness functions to encourage different types of behavior. Scenario specifics are shown below each of the following videos.
Predator-Prey Scenario: Demonstrates the acquisition of pursuit and evasion strategies. Here, a small group of slow predatory robots evolved to capture individuals from a fast group of prey robots. Prey robots evolved strategies of survival.
Setup: The environment was divided into three sections. Two predator robots were placed on one side of the map and six prey robots were placed at the opposite side. For this scenario, prey robots were not destroyed if they crashed into each other and predator robots were unaffected by any robot. Robots were still required to avoid walls.
Fitness Evaluation: Predator teams were encouraged to quickly seek out and destroy prey robots while prey robots were simply encouraged to survive. Predator robots that intercepted prey robots were rewarded with an increase in fitness that was proportional to the amount of time left in that trial. Prey robots were assigned a fitness that was equal to the number of iterations they survived. All robots received a fitness of zero if they crashed into a wall.
Quantitative Evaluation: The fitness plot is shown in the background of the video. Each predator success is shown as a green line while each prey is shown as a red line. After the initial generations, the learning curves for both predators and prey remained relatively flat. This may be due to the fact that both predators and prey were getting better at their job at about the same rate.
Qualitative Evaluation: It didn’t take long for the prey robots to discover that they were prey. After only a few generations, they were starting to show dodging and fleeting behaviors. It is unclear when (or if) the predators learned that they were predators. Many just appeared to drive around at top speeds. In later generations, some predators were found to reverse direction if a prey was following them and many would risk getting close to a wall if a prey was trapped there. Predators also appeared to form a strategy in the looped hallway, where one predator would take the counter-clockwise route, while the other would proceed in a clockwise direction. This would allow prey to be trapped and captured in the center.
Mutualism Scenario: Demonstrates cooperation. Two dissimilar groups of robots evolved the ability to work together to accumulate points.
Setup: The environment was divided into three sections. Two grouper robots were placed on one side of the map and six groupie robots were placed at the opposite side. For this scenario, robots were not destroyed if they crashed into each other but they were still required to avoid walls.
Fitness Evaluation: Groupers were encouraged to have groupie robots following them. Groupie robots were encouraged to follow grouper robots. If the distance between the following robot and followed robot was between 2 and 6 meters and if the groupie was facing the grouper, both robots received a point for the iteration. Grouper robots could receive multiple points in an iteration if they were pursued by multiple groupie robots.
Quantitative Evaluation: Groupers and Groupies steadily increased their fitness throughout training.
Qualitative Evaluation: In early iterations, groupies were most obvious. Within 20 iterations, groupies were regularly observed attempting to pursue their groupers, though they often failed at the task. Over dozens of iterations, the groupers appeared to be assisting the groupies by reducing their speed. By the hundredth iteration, robots behaved very productively. Groupies formed nice tight clusters behind groupers and groupers would drive slowly and maneuver in ways that would facilitate the collection of groupies. Groupers also learned that they could acquire the most points if they stuck together. Even if a grouper possessed all the groupies, it would still speed up to catch the other grouper.
Parasite-Host Scenario: Demonstrates pursuit and evasion. Hosts must develop strategies to avoid fast-moving parasites.
Setup: The environment was divided into three sections. Two host robots were placed on one side of the map and six parasite robots were placed at the opposite side. For this scenario, robots were not destroyed if they crashed into each other but they were still required to avoid walls.
Fitness Evaluation: Hosts were encouraged to evade the parasites. Parasites were encouraged to remain in the proximity of hosts. If the distance between the parasite and host was less than 3 meters, the parasite received a point for the iteration. The host robot automatically received a point on every iteration but lost a sixth of a point for every parasite it had acquired for that iteration.
Quantitative Evaluation: After about 40 generations of evolution, hosts lost their advantage over the parasites, after which, the learning curve remained relatively flat.
Qualitative Evaluation: The robots used in these experiments have a drive setting for slow speeds, but not one to allow them to stop. Early in the training, hosts exploited this fact and were able to loose many of their parasites by entering into a spin behavior. This behavior essentially halted the forward movement of the host and parasites continued past. It didn’t take long before parasites adapted to this behavior. In response, parasites started displaying a circling behavior around the host. When this happened, hosts completely stopped with the spin behavior and began a new behavior, where they would move at minimum speed until an infection occurred, after which, the robot would speed up and would make apparent attempts to maneuver the parasites onto its teammate.