Effects of Personality Distribution on Collective Behavior

Optimizing group success is challenging for multi-robot systems, especially for large systems such as robot swarms where even simple individual interaction rules can lead to complex group behavior. Studies of natural systems have shown that heterogeneous groups can outperform homogeneous groups, especially when individual differences lead to role or niche specialization. This happens even when the individuals are seemingly identical. Although individuals within a group may appear physically identical, they can vary widely in their personality, which significantly affects their behavior. However, determining the most effective composition of personalities in a group is particularly difficult in natural systems given the ambiguities of animal personalities and the physical challenges of repeated evaluations. Using a biologicallybased collective movement model, we evaluate different personality distributions to determine their effect on the overall success of the group. Results show that although there are distributions that are clearly more effective than others, in many cases, there is a broad range of distributions that results in high group success. Furthermore, experiments using variable, or adaptive, personalities demonstrate that successful distributions are stable equilibriums as initially extreme distributions converge to these successful distributions as personalities change. Introduction In many robot swarms, individuals within the swarm are assumed to be heterogeneous, especially, and perhaps most importantly, with respect to their decision-making. In fact, heterogeneity is argued to be a requirement for a multi-robot system even to be considered a swarm (Şahin, 2005). Variations in either morphology or decision-making can present significant challenges given the complexities that arise as robots interact. However, such differences can also provide significant benefits (Couzin et al., 2005). While most work involving these benefits focuses on morphological differences, of particular interest here are differences in the decision-making process, especially those that are adaptive. The ability for individuals in a group to adapt their decision-making to improve group-level performance is key to the usefulness of multi-robot systems in dynamic, real-world tasks. Recent research in animal groups has shown that such systems exist in nature. Animals are known to exhibit sets of correlated traits referred to as personalities, that in part, determine their behavior (Sih et al., 2004). Furthermore, these personalities are known to change in response to both individual experiences and the distribution of personalities in the overall group (Dall et al., 2004; Montiglio et al., 2013). As such, animal personalities can provide inspiration as a means of integrating adaptive individual differences within a group to improve performance. Furthermore, we have previously shown that using a simple mechanism to introduce personality into a collective movement model can result in a significant increase in success (Eskridge et al., 2013). However, it is not well understood how the distribution of personalities affects group performance, nor how stable effective personality distributions are. In the work presented here, we use a biologically-based collective movement model to explore how distribution of personalities, especially as they pertain to leaders and followers, affects the overall success of the group in initiating a movement. Simulations demonstrate that, for many of the groups evaluated, there is a wide range of personality distributions that results in high group success. Furthermore, the results demonstrate that these successful distributions are stable equilibriums. This is indicated not only by the wide range of successful distributions, but also by additional experiments using variable personalities in which extreme personality distributions converge to the more successful distributions as personalities change.