Scientists have long wondered whether the schooling of fish is driven by hydrodynamic benefits, but the complexity of their environment makes unraveling this complex motion difficult. A recent study uses a different tactic, combining direct numerical simulation of the fluid dynamics with techniques from artificial intelligence and machine learning to build and train autonomous, smart swimmers.
The authors use a technique called deep reinforcement learning to train the swimmers. Essentially, the swimmer being trained is able to observe a few variables, like its relative position to the lead swimmer and what its own last several actions have been – similar to the observations a real fish could make. During training, the lead swimmer keeps a steady pace and position, and the follower, through trial and error, learns how to follow the leader in such a way that it maximizes its reward. That reward is set by the researchers; in this case, one set of fish was rewarded for keeping a set distance from their leader, one intended to keep them in a position that was usually beneficial hydrodynamically. Another set of fish was rewarded for finding the most energy-efficient method for following.
Once trained, the smart swimmers were set loose behind a leader able to make random decisions. Above you can see the efficiency-seeker chasing this leader. Impressively, even though this smart swimmer had the option to go it alone (and had never followed such a dynamic leader), it does an excellent job of keeping to the leader’s wake. Compare it with real swimmers and there’s a definite similarity in their behavior, which suggests the technique may be capturing some of an actual fish’s intuition. (Image and research credit: S. Verma et al., source; thanks to Mark W. for assistance)
