Integrating computational science with biology to study collective animal behaviorRandal S. Olson and Christoph Adami

Michigan State University, East Lansing, Michigan

Biologists have spent decades studying collective animal behavior due to its important implications forsocial intelligence, collective cognition, and potential applications in engineering, artificial intelligence, androbotics [1]. Of the numerous forms of collective animal behavior, swarming behavior stands as one of themost striking examples observed in nature. Since the long generation times in swarming animals makesstudying the evolution of swarming behavior difficult [3], we developed a computational model that simu-lates digital organisms with evolving behaviors. Using this model, we determined that the confusion effect,where swarming prey confuse and thereby reduce the attack efficiency of their predators, provides a suffi-cient selective advantage to evolve and maintain swarming behavior in prey in a digital model [6]. However,many aspects of the importance of predator confusion for the evolution of swarming remain unanswered.

Recently, Ioannou et al. (2012) designed an innovative study where hand-coded, simulated prey wereprojected onto the side of a fish tank containing a single predatory fish, allowing a highly controlled study ofwhat aspects of prey swarming behavior affect predator hunting behavior [2]. While Ioannou et al. (2012)demonstrated that swarming behavior can be selected for by predation alone once swarming behavior ispresent in the population, the question of how swarming behavior arose in the population in the first placehas yet to be explored. Here we propose to extend this experiment by allowing the simulated prey behaviorsto evolve in response to selective pressures applied by the biological predator. This will enable us to addresshypotheses about swarm behaviors in response to biological predators on an evolutionary scale, as opposedto studying swarming behavior at a fixed point in evolutionary time.

Research Hypotheses, Methods, and Expected Outcomes Using our computational model of Darwinianevolution (Figure 1), we will examine the predator-prey dynamics between simulated swarming water fleas(Daphnia magna) and predatory three-spined sticklebacks (Gasterosteus aculeatus) to address the hypothe-ses outlined in this section. All proposed experiments will be repeated 30 times (each with a differentstickleback to avoid the effects of learning), using video analysis to measure the number of capture attemptsby the stickleback on the simulated prey and the latency to the first attack by the stickleback on the simulatedprey. We will compare the mean attack efficiency (# successful attacks / total # attacks) and time to firstattack attempt to detect if there are significant differences in the predator’s response to the experimental preybehaviors.

Hypothesis I: Swarming behavior decreases predator attack efficiency. Previous work has suggested thatDaphnia swarming behavior decreases the attack efficiency of predatory three-spined sticklebacks [5]. Wewill seek to confirm this hypothesis in our system by projecting groups of 50 simulated prey with pre-evolved behaviors onto the side of a fish tank containing a single stickleback. We expect prey that swarm toexperience fewer successful attack attempts from the predator than prey that move around randomly.

Hypothesis II: Larger swarms reduce predator attack efficiency more than smaller swarms. Next, wewill repeat the first experiment with pre-evolved cohesive swarms and vary the number of simulated preyin the swarm (swarm sizes from [3]: 5, 15, 25, 50, 100). This hypothesis predicts that the sticklebacks willperform fewer successful attacks and take longer to attack larger swarms (size 25, 50, and 100) than smallerswarms (size 5 and 15). If this prediction holds, it would indicate that larger swarms increase the difficultyof predator attacks on individual prey, which is likely the result of the confusion effect. Alternatively, ifthere is no difference in predator response between experiments, then this would suggest that predators thatfeed on swarming prey are not affected by the confusion effect. Lastly, if predators instead prefer to attacklarger swarms than smaller swarms, then this would suggest that the confusion effect is not magnified byswarm size, and attacking larger swarms is advantageous for predators because there is potentially moreprey to be captured per attack.

1

Figure 1: The proposed integrated experimental system, where the interaction of real predators with sim-ulated swarms will be tested. A population of 30 prey Markov Networks (MNs) will be evolved usinga genetic algorithm (GA). For every GA generation, each MN will control a group of 50 simulated preyagents for a 30-minute trial period. During the trial period, camera-recorded predator attacks on the groupof simulated prey affect the fitness of its MN, so that prey can be subjected to realistic rather than simulatedpredation. After the evaluation period, each MN will be assigned a fitness inversely proportional to thenumber of attacks the MN’s group of prey experienced. Thus, the more an MN’s group of prey is attackedduring the trial period, the less likely the MN is to produce an offspring into the next GA generation.

Hypothesis III: The confusion effect provides a sufficient selective advantage to evolve swarming behav-ior in simulated prey. Once we have established that sticklebacks are confused by simulated prey swarmingbehavior, we can then test the key discovery in our previous work [6]. We will repeat the first experimentwith initially random prey behaviors, except the prey behaviors will be allowed to evolve in response to thepredator. Following the evolutionary process described in Figure 1 (more details in [6]), we expect prey toevolve increasingly cohesive swarming behavior in response to the predator over evolutionary time. If theprey evolve cohesive swarming behavior, then this experiment will provide strong evidence that predatorconfusion selects for the evolution of swarming behavior in prey. Conversely, if the prey do not evolve co-hesive swarming behavior, then this experiment will highlight the need for further exploration into the otherpossible selective advantages that could select for the evolution of prey swarming behavior [4].

Hypothesis IV: The structure of the predator’s visual system plays a significant role in the efficacy ofthe confusion effect. Lastly, we will test the hypothesis proposed in our previous work that the efficacyof the confusion effect as a defensive mechanism can be reduced if the predator evolves a more focusedvisual system [6]. The stickleback has two species variations that exhibit significantly different foragingbehavior: limnetics typically feed on plankton in clear water near lake surfaces, whereas benthics feed onsmall invertebrates in the cloudy water on the lake floor [7]. As in the previous experiments, we will project50 simulated prey onto the side of fish tanks containing separate species. This hypothesis predicts that thelimnetics will exhibit significantly higher attack efficiency and shorter attack latencies than the benthics dueto the limnetics’ specialized, focused visual system for hunting agile prey in clear water. If this predictionholds, then this suggests predators that feed on swarming prey could have a selective advantage by evolvinga narrow, focused retina to reduce the efficacy of the confusion effect.

2

Intellectual Merit This interdisciplinary research advances the field of behavioral science by merging abiological system with an evolving computational system, offering researchers unprecedented experi-mental control over predator-prey dynamics and the ability to directly test hypotheses about the evolu-tion of behavior in response to predation. Michigan State University offers the necessary facilities for thisresearch, including Dr. Jenny Boughman’s research lab to perform the experiments with live sticklebacks;Dr. Christoph Adami’s computational lab studying evolutionary processes and the evolution of animal be-havior to prepare the evolving digital swarm platform; guidance from behavioral biologist Dr. Fred Dyer tointerpret the sticklebacks’ behavior; and support from the NSF BEACON Center, an institution promotinginterdisciplinary research collaborations between biologists, computer scientists, and engineers.

Broader Impacts This work develops a platform for directly interfacing biological and computationalresearch, and increasing our understanding of collective cognition and decision-making in animals. Theavailability of such an interface should open up a new frontier in the study of the evolution of animalbehavior and artificial intelligence. In addition, research in digital evolution and animal behavior is readilyaccessible to broad populations, and we will continue to share this research with college undergraduates andK-12 audiences through blog and video blog posts on our lab web site. Further, after we establish that preywill evolve swarming behavior in response to predation from live fish, we will create a simplified versionof this platform which enables students to “predate” on the simulated prey populations via a touchscreeninterface and observe the evolution of prey behavior in response to their predation strategies. Thus, theproposed platform will provide a powerful tool for scientific endeavors as well as outreach and teachingefforts.

References[1] ID Couzin. Collective cognition in animal groups. Trends in Cognitive Sciences, 13(1):36–43, 2009.

[2] CC Ioannou, V Guttal, and ID Couzin. Predatory fish select for coordinated collective motion in virtualprey. Science, 337(6099):1212–1215, 2012.

[3] JM Jeschke and R Tollrian. Prey swarming: which predators become confused and why? AnimalBehaviour, 74(3):387–393, 2007.

[4] J Krause and GD Ruxton. Living in Groups. Oxford University Press, USA, 2002.

[5] M Milinski and R Heller. Influence of a predator on the optimal foraging behaviour of sticklebacks(gasterosteus aculeatus l.). Nature, 275:642–644, 1978.

[6] RS Olson, A Hintze, FC Dyer, DB Knoester, , and C Adami. Predator confusion is sufficient to evolveswarming behavior. Royal Society Interface, in review, 2013.

[7] S Ostlund-Nilsson, I Mayer, and FA Huntingford. Biology of the Three-Spined Stickleback. Taylor &Francis, 2006.

3