Explaining Invasive Ant Scenarios Using Emotion Modeling Joshua Wainer, Dusan Jan, and MinHee Kwon Introduction Native ant species play important roles in the stability of their ecosystems, acting as sources of food for many animals, natural predators for some, and facilitators in the life cycles of many plants. When native ant species are displaced or killed off by an invasive ant species such as Linepithema humile (Argentine ants), the latter rarely fills the same biological niches or assumes the same ecological roles as the former. Subsequently, ecosystems that were once stable are usually made unstable by the arrival of an invasive ant species [7]. While the phenomenon of an invading species disrupting a local ecosystem is as old as biology itself, the exact mechanisms through which such upheaval takes place are rarely understood. However, recent research [2] has helped to explain the processes and interactions behind the successful invasions of L. humile into areas with Mediterranean climates, such as California. Because this area owes much of its economic success to its agriculture which in turn is dependent on the stability of local ecosystems, a proper understanding of the processes that drive invasive ant behaviors is critical. To this end, we will attempt to use a simple computational model of emotions to synthesize both the local interactions among invasive and native ant species as well as the overall displacement of the latter by the former. Related work There are many papers that have dealt with interspecific ant interactions. Holway found that invasive Argentine ants exhibit fewer intraspecific conflicts than their native counterparts which leads to higher population densities, more foraging activity, and greater brood production, which presumably contribute to their successful invasions [8]. Later, he found while Argentine ants used chemical compounds as well as physical aggression when fighting with native ants, the key to their successes in fights lay in their numerical superiority. Furthermore, Argentine ants were unique in that they were both good fighters as well as good foragers, while most native ants had a negative correlation between the two traits [9]. Davidson had noted the same phenomenon in other invasive species and found that this could be attributed to factors such as using carbohydrates to replace nitrogen in their diets, the practices of polydomy and polygyny, and/or the formation of supercolonies [13]. Human and Gordon were the first to suggest importance in the Argentine ant’s ability to balance foraging success with fighting success [10], but their later work suggests that aggressive behavior does not correlate at all with foraging success among any ant species. Instead, it supports the idea that success at displacing other ant species was positively correlated with a high frequency of initiating interspecific interactions [2]. Additionally, there have been many papers that have attempted to recreate or mathematically model various behaviors of ants. Sumpter and Beekman showed that an ant’s response to a pheromone trail is nonlinear with respect to the amount of pheromone deposited and seems to suggest an activation threshold in ants for following a trail. Additionally, the amount of pheromone that ants laid on a trail directly correlated with the level of quality of the food source, and ants tended to follow a trail that had the most ants following it [11]. Chialvo and Millonas also studied the probabilities of ants moving in random directions as well as following a given pheromone trail and found that the empirically-determined probabilities enabled the formation of “well-defined” trails [14]. Work has also been done that shows the validity of using emotions to drive the behavior of agents. Scheutz showed the validity of using emotions such as “anger” and “fear” as driving factors behind the actions of many agents [12] as well as how displaying aggression is beneficial for social groups of agents[15]. Methods To test whether an emotional model of ant behavior could synthesize the ants’ interspecific interactions and the overall result of an invasion by Argentine ants, we constructed a simulated environment in which two species of ants had their nests randomly placed within a 2D world. With all the ants of a given species starting at their nests, they began foraging for food at the same time. At random intervals, user-specified chunks of food would appear somewhere in the ants’ world. When the ants found food, they would try to take it back to their nest and store it, but if they had an interaction with another ant of a different species, there was a possibility that this would not happen. Operating under the assumption that a species of ant would either die out or be displaced in order to find food elsewhere when they gathered significantly less food than another species living in the same environment, we monitored how much food each species gathered in a given simulation. Our simulation recreated both the behaviors common to all ants as well as the interspecific interactions found between various ant species. The behaviors equally prevalent in all species included randomly foraging, following a trail of pheromones to food, and returning with food to the nest. These behaviors were modeled as a set of states in every simulated ant, with foraging for food set as a default behavior. If our simulated ants came across a pheromone trail laid by their own species, they would follow such a trail away from their nests since it would be laid down with the intention of leading ants to food [1]. Once the ants found food, they would take it directly back to their nests while laying down a trail of pheromones for other ants of their own species to follow. While the above behaviors occur without any communication at all among the ants, there is a specific set of behaviors that take place only when two or more ants interact with each other. Such behaviors include biting, lunging, running, and antennating (communicating through the touching of antennae), and can be grouped into three categories of behavior: aggressive, retreating, and neutral. These behaviors are found with varying degrees of frequency among different species of ants [2], though there are protocols of interaction that all ants seem to comply with: Anger Fear ants will never run away when they begin an interaction with another ant, but ants will usually run away when approached by another ant. ants will respond with aggressive behavior if and only if such behavior is first displayed toward them ants will begin an interaction with aggression more often when they are close to a large number of ants of their own species [3] Stimulus Not acting aggressively Interacting with an ant of a different species, regardless of who initiated the interaction Decay Time, acting aggressively Time, co-location with ants of same species Table 1: Operational Laws of Ants' Emotions To drive these behaviors, we used a simple computational model of emotions originally conceptualized by Sloman [4] and later developed by Velasquez. In this model, an emotion is simply a value that changes over time due to stimulation (increasing value) and decay (decreasing value). Emotions also have saturation limits to ensure that they do not exceed a maximum value as well as activation thresholds that determine when a given emotion causes a particular behavior to activate [5]. Considering that there is no correlation between the frequency of displayed aggression by a species and the outcome of interactions between ants [2], we separated interspecific interactions into two components: to willingness to begin an interaction, and the willingness to act aggressively. We decided to call the emotions driving such behaviors “fear” and “anger”, respectively, though it is important to note that the dynamics behind these emotions are not meant to be similar to the mechanisms behind the human emotions of the same names. The operational laws behind the two emotions are described in Table 1. It is important to note that while the decay law for fear is drawn from scientific literature [3], the other laws were made to provide stability for the emotional values. In addition to modeling the ants’ internal behaviors and drives, special attention was paid to environmental variables that could affect the outcome of the simulation, such as the location of the random food drops, their frequency, the size of the world, the number of ants simulated, and the duration of a given simulation. Specifically, if food was dropped closer to one species’ nest than another, this would give the closer species an unfair advantage in gathering a large amount of food, thus skewing our data. Additionally, if food were dropped too frequently, the ants were very likely to go for food source closest to them and were unlikely to interact with the other species at all, skewing our data again. Furthermore, if too few or too many ants were placed in a world of a particular size, this would result in findings that were not gathered under the same conditions as those in the study whose results we were attempting to recreate. In short, because any of the above circumstances would invalidate our findings, we had to take extra care in tuning these variables. The abovementioned rules, algorithms, and techniques were all implemented in Swarm, an open-source toolkit used for simulating the individual behaviors of many independent agents to generate emergent group behaviors [6]. The toolkit features two separate modes of operation, visual and batch, which allow for faster, easier debugging and impressive demonstrative capabilities as well as decreased running time for simulations, respectively. Additionally, the toolkit offers the choice of programming in either Java or Objective-C, a variant of the C language developed by Apple. We opted for the latter choice, as there are more programs that demonstrate the full capabilities of Swarm in Objective-C. Use of experiments in tuning the model After we had the basic framework of the simulation in place we had to fit the model to the data reported in field observations. We had to ensure that that the model generalizes well for parameters outside the ones observed in the field and that the match for the observed ones was as close as possible. In the experiments conducted by Human and Gordon the maximum numbers of ants involved in interactions at baits were about 20 to 50 of each species. With this in mind we decided to test our model for 25, 50, 75 and 100 ants of each species. In the initial tuning we focused on a specific pair of ants, the Argentine ants Linepithema Humile (LH) and one of the native species Messor Andrei (MA). The reports by Human and Gordon indicate that LH initiated 75% of all interactions. Of the interactions initiated by LH, 90% of them were initiated aggressively. MA on the other hand initiated aggressively only 47% of the time. The first factor we tried to control was the ratio of initiated interactions. Our initial model increased fear on interaction with other species, but only decayed it based on time. A simple mathematical model indicated that using a ratio of 3:1 for fear decay of HA and MA should give the desired effect, but it was unclear how reliable it was as many parameters depended on interactions between the two species. After tuning the decay parameters for 25 ants per species we checked how well the behavior generalized with higher numbers. We found that the ratio was out of bounds at higher populations. The problem was in the fact that fear decay was the same in all scenarios while fear increase scaled with number of interactions. This coupled with a relatively high saturation threshold used for fear put MA in constant state of fear. To counterbalance this we had to create a decay mechanism that would scale as the population increases. We decided to reduce fear with interactions with ants of same species. While this achieved the desired property of having decay scale with population size it is also supported by literature which suggests that ants that are alone and away from their colony are less prone to interact with other ants. Using this additional decay mechanism and lower saturation limit for fear (twice the value of activation threshold) we achieved results that are much more even across different population sizes. While there were still some variations it was nowhere close to the spread we observed before (before ratios ranged up to several hundred). Table 2 shows the number of initiated interactions for each species and the ratio for each of the 4 colony sizes. While the ratio did not fit exactly to the desired 3:1 we left the fine tuning until after other aspects of the simulation were finalized. 25 ants per species LH MA Ratio 993 670 1042 641 1096 746 1461 567 953 795 830 700 836 613 583 465 1050 797 917 479 Average Deviation 976.1 225.86 647.3 118.55 50 ants per species LH MA Ratio 3802 1731 2844 1523 3852 1294 2479 1915 2686 1710 2647 1700 2805 1595 2530 1857 2350 1706 2746 1873 1.48 1.63 1.47 2.58 1.2 1.19 1.36 1.25 1.32 1.91 1.54 0.43 2874.1 524.46 1690.4 185.04 2.2 1.87 2.98 1.29 1.57 1.56 1.76 1.36 1.38 1.47 75 ants per species LH MA Ratio 5748 2860 7099 3114 4897 3029 6693 3410 7246 2596 6440 3557 6732 3085 6169 2998 5456 2931 5277 2679 1.74 0.51 6175.7 801.99 3025.9 295.02 2.01 2.28 1.62 1.96 2.79 1.81 2.18 2.06 1.86 1.97 2.05 0.32 100 ants per species LH MA Ratio 10461 4505 9185 4665 8706 4991 8254 4598 11971 3904 10386 5520 13102 3941 9331 4942 9853 4530 10010 3995 10125.9 1474.48 4559.1 516.66 2.32 1.97 1.74 1.8 3.07 1.88 3.32 1.89 2.18 2.51 2.27 0.55 Table 2: Number of initiated interaction between HA and MA in 10000 cycles The tuning of anger was easier as the value was influenced mainly by decisions of the ant itself, so a mathematical model was more accurate. Having an anger increase of A when initiating interactions in a neutral manner, r being the desired ratio of aggressively initiated interactions, t being average time between encounters initiated and T total elapsed time then total anger increase is A*T/t*(1-r) and total anger decay is T. These two have to be balanced so the anger increase should be set to A=t/(1-r) to achieve the desired ratio. The variations resulting from changes in anger from reactive behavior also have some minor impact, but experiments confirmed that setting the anger increase in this way achieved the desired ratio and was quite stable. To measure which species was more successful we looked at ratio of food collected for each species. Since the total amount of food could vary between simulations the difference was not the right measure. Also to get a linear measure for the result variable we applied log to the ratio of food collected. A result of 0 means that both species collected the same amount of food, positive values indicate HA collected more food and negative values indicate MA collected more food. Looking at the collected simulation data it was clear that there was no significant difference between foraging success of the two species. At this point we started refining the simulation in order to minimize the effects on foraging success by variables that we were not interested in. By looking at the visual runs of the simulation it seemed that the most significant factor influencing foraging success was distance of food drops to the nest. To eliminate this bias we After we had the desired local behaviors in place we started collecting data on foraging success of the two species. Table 3 shows the food collected after 10000 cycles in the first iteration of the simulation design. Average Deviation 25 ants per species HA MA Result 2.66E+03 3.73E+03 -1.48E-01 3.89E+03 1.69E+03 3.63E-01 2.59E+03 1.97E+03 1.19E-01 1.35E+03 2.18E+03 -2.10E-01 2.74E+03 1.83E+03 1.76E-01 3.76E+03 2.90E+03 1.13E-01 1.01E+03 4.54E+03 -6.53E-01 3.29E+03 3.48E+03 -2.45E-02 1.30E+03 1.89E+03 -1.61E-01 2.89E+03 2.62E+03 4.22E-02 2.55E+03 2.68E+03 -3.82E-02 1019.64 963.38 0.28 Average Deviation 50 ants per species HA MA Result 3.73E+03 4.57E+03 -0.09 5.51E+03 3.30E+03 0.22 3.29E+03 7.93E+03 -0.38 3.23E+03 5.55E+03 -0.23 2.87E+03 3.47E+03 -0.08 3.94E+03 4.77E+03 -0.08 6.61E+03 2.45E+03 0.43 5.38E+03 2.99E+03 0.26 3.55E+03 7.19E+03 -0.31 2.95E+03 4.28E+03 -0.16 4.11E+03 4.65E+03 -4.31E-02 1277.08 1795.39 0.26 Average Deviation 100 ants per species HA MA Result 4.19E+03 6.09E+03 -0.16 5.78E+03 4.94E+03 0.07 6.78E+03 5.34E+03 0.1 7.26E+03 9.21E+03 -0.1 6.69E+03 4.83E+03 0.14 7.03E+03 5.54E+03 0.1 6.92E+03 3.58E+03 0.29 6.72E+03 3.34E+03 0.3 3.78E+03 7.98E+03 -0.32 9.87E+03 5.01E+03 0.29 6.63E+03 5.65E+03 0.07 8.51E+03 2.97E+03 0.46 7.24E+03 4.39E+03 0.22 5.87E+03 7.95E+03 -0.13 9.05E+03 5.69E+03 0.2 4.71E+03 4.79E+03 -0.01 9.00E+03 6.03E+03 0.17 8.29E+03 4.49E+03 0.27 6.47E+03 6.79E+03 -0.02 4.08E+03 3.33E+03 0.09 6.74E+003 5.40E+003 0.1 1695.42 1644.07 0.19 Table 3: Food collected with random food dropping Average Deviation 25 ants per species LH MA Result 2.33E+03 6.92E+02 0.53 2.42E+03 3.23E+03 -0.13 1.60E+03 1.66E+03 -0.02 2.29E+03 2.24E+03 0.01 1.34E+03 2.05E+03 -0.19 1.41E+03 1.33E+03 0.03 1.11E+03 7.20E+02 0.19 1.32E+03 1.71E+03 -0.11 1.90E+03 1.69E+03 0.05 2.17E+03 2.77E+03 -0.11 0.03 0.21 50 ants per species LH MA Result 5.91E+03 4.71E+03 0.1 2.82E+03 4.39E+03 -0.19 2.61E+03 2.55E+03 0.01 2.83E+03 1.96E+03 0.16 2.55E+03 3.70E+03 -0.16 4.77E+03 4.00E+03 0.08 2.67E+03 3.47E+03 -0.11 3.20E+03 2.73E+03 0.07 2.69E+03 2.74E+03 -0.01 4.60E+03 3.95E+03 0.07 0 0.12 75 ants per species LH MA Result 4.01E+03 4.26E+03 -0.03 4.68E+03 3.60E+03 0.11 4.17E+03 5.14E+03 -0.09 6.17E+03 6.15E+03 0 5.14E+03 4.87E+03 0.02 4.87E+03 6.16E+03 -0.1 4.75E+03 3.09E+03 0.19 4.45E+03 3.35E+03 0.12 6.03E+03 5.16E+03 0.07 4.32E+03 3.23E+03 0.13 0.04 0.1 100 ants per species LH MA Result 3.77E+03 4.51E+03 -0.08 5.99E+03 3.77E+03 0.2 4.56E+03 5.71E+03 -0.1 2.90E+03 2.28E+03 0.11 3.91E+03 4.50E+03 -0.06 4.57E+03 3.97E+03 0.06 5.66E+03 4.47E+03 0.1 7.14E+03 7.27E+03 -0.01 6.73E+03 5.73E+03 0.07 4.16E+03 4.23E+03 -0.01 0.03 0.1 Table 4: Food collected with equidistant food locations made all food drops appear at locations equidistant to nests of both species. Table 4 shows the results after this change. While the deviation reduced by this change, the results were still chancy (paired t-test of log values not significant). Careful examination of the simulations revealed that most of the interactions that happened were by chance and in most cases the species collected the food from different sides of the food patch, rarely interacting until the food was almost gone. Since literature reported that ants patrol the areas around food locations we decided to implement this behavior as we believed it would have a significant impact on number of interactions between the two species at the foraging sites. We examined several possible algorithms for patrolling and we finally decided to use a model where an ant randomly chooses a direction where there is food present and walks in that direction for 2 simulation cycles. The algorithm had a nice property that the ants walked randomly along the food patch, but were not stopped by small gaps in food location. The ants would patrol the area for 5 cycles before returning to nest with the food. Table 5 shows the results after we implemented the patrolling behavior. Average Deviation P value 25 ants per species LH MA Result 2.03E+03 2.08E+03 -0.01 1.63E+03 5.86E+02 0.44 3.11E+03 1.92E+03 0.21 4.62E+02 1.75E+03 -0.58 2.22E+03 1.16E+03 0.28 4.19E+03 1.10E+03 0.58 1.08E+03 1.51E+03 -0.15 2.62E+03 1.19E+03 0.34 7.34E+02 1.58E+03 -0.33 1.95E+03 1.96E+03 0 1.53E+03 1.10E+03 0.14 2.05E+03 9.98E+02 0.31 1.70E+03 2.09E+03 -0.09 1.24E+03 1.07E+03 0.06 1.27E+03 1.64E+03 -0.11 1.78E+03 1.46E+03 0.09 1.64E+03 1.28E+03 0.11 1.34E+03 1.66E+03 -0.09 1.99E+03 9.97E+02 0.3 1.68E+03 2.95E+03 -0.24 0.06 0.28 0.3240 50 ants per species LH MA Result 2.83E+03 1.86E+03 0.18 2.31E+03 2.88E+03 -0.1 2.07E+03 3.38E+03 -0.21 3.44E+03 2.70E+03 0.11 3.63E+03 1.91E+03 0.28 4.63E+03 1.87E+03 0.39 2.56E+03 3.78E+03 -0.17 4.41E+03 2.90E+03 0.18 1.56E+03 1.39E+03 0.05 4.65E+03 3.06E+03 0.18 3.76E+03 3.44E+03 0.04 2.65E+03 2.01E+03 0.12 1.37E+03 1.44E+03 -0.02 3.23E+03 1.96E+03 0.22 3.11E+03 2.10E+03 0.17 3.11E+03 1.41E+03 0.34 3.71E+03 3.21E+03 0.06 3.08E+03 1.73E+03 0.25 4.36E+03 2.10E+03 0.32 4.26E+03 1.83E+03 0.37 0.14 0.17 0.0020 At this point the paired t-tests showed significant results for all tests except in the 25 ant case. Since we introduced some changes to the simulation we went back and retuned the fear and anger values. We achieved ratios of initiated interactions of 3.3 for 100 ants, 2.9 for 75 ants, 2.4 for 50 ants ant 1.7 for 25 ants. Our observations suggest that main difference in ratios came from the fact that at lower population sizes there are more interactions by chance and less interactions at foraging sites. Most interactions by chance occur when both ants have low fear value which makes them equally likely to initiate the interactions, thus lowering the ratio for lower population sizes. It would be interesting to compare ratios for interactions at foraging sites only, but we did not perform such an analysis in our study. After we had conclusive results for interaction between LH and MA we tuned the parameters for another native species, in this case Camponotus Semitestaceus (CS) which is more aggressive than MA (79% of initiated interactions were aggressive). In simulations of LH and CS the results were significantly in favor of LH (p = 0.0001), which shows that aggression 75 ants per species LH MA Result 4.89E+03 4.07E+03 0.08 6.57E+03 3.19E+03 0.31 4.26E+03 3.61E+03 0.07 3.84E+03 3.30E+03 0.07 5.03E+03 3.96E+03 0.1 4.77E+03 2.99E+03 0.2 3.79E+03 4.10E+03 -0.03 5.70E+03 3.67E+03 0.19 6.00E+03 3.81E+03 0.2 3.31E+03 3.62E+03 -0.04 8.90E+03 2.61E+03 0.53 5.67E+03 5.33E+03 0.03 4.13E+03 3.37E+03 0.09 2.75E+03 3.42E+03 -0.09 5.17E+03 3.43E+03 0.18 3.70E+03 4.48E+02 0.92 5.68E+03 3.55E+03 0.2 2.38E+03 2.69E+03 -0.05 7.35E+03 2.18E+03 0.53 4.49E+03 3.29E+03 0.14 0.18 0.24 0.0033 Table 5: Food collected with patrolling behavior 100 ants per species LH MA Result 8.99E+03 2.67E+03 0.53 8.35E+03 1.94E+03 0.63 5.45E+03 1.92E+03 0.45 5.53E+03 2.97E+03 0.27 6.95E+03 3.49E+03 0.3 5.45E+03 2.94E+03 0.27 6.66E+03 3.25E+03 0.31 8.34E+03 4.42E+03 0.28 4.72E+03 4.56E+03 0.02 4.20E+03 4.71E+03 -0.05 6.75E+03 2.51E+03 0.43 9.20E+03 3.41E+03 0.43 5.95E+03 4.01E+03 0.17 4.35E+03 4.68E+03 -0.03 6.06E+03 2.66E+03 0.36 7.11E+03 4.09E+03 0.24 6.68E+03 4.20E+03 0.2 4.82E+03 3.46E+03 0.14 4.41E+03 2.37E+03 0.27 6.54E+03 3.95E+03 0.22 0.27 0.18 0.0001 levels do not play a big role in foraging success as indicated in the literature. To verify this we conducted another set of simulations between MA and CS which only differed in aggression. The results again indicated that aggresion does not have a significant impact on foraging success (p = 0.8634). Conclusions In this paper we presented an emotion-inspired model for simulating local ant behavior and showed that computer simulations confirm the hypothesis presented by biologists. Comparisons of foraging success in simulations between ants of different species show that foraging success is not related with frequency of aggression, but is largely dependent on who initiates the interactions. While the Argentine ants are also numerically dominant in the intoduced areas, these behavioral tendencies affect their success as an invader to a greater degree. The fact that their numbers are high further helps them to explore space more effectively as well as encounter and displace ants of native species more frequently at food sources. References [1] Carthy, J.D. The orientation of two allied species of British ant. 11. Odour trail laying and following in Acanthomyops (Lasius) fuliginosus. Behaviour, 3 : 304-318, 1951 [2] Human, KG and Gordon, DM. Behavioral interactions of the invasive Argentine ant with native ant species. Insectes Sociaux, volume 46, no. 2, p. 159--163,1999 [3] Gordon, D.M.. How colony growth affects forager intrusion in neighboring harvester ant colonies. Behav. Ecol. Sociobiol. 31: 417–427. 1992 [4] A. Sloman and M. Croucher. Why robots will have emotions. In Proceedings of the Seventh International Joint Conference on Artificial Intelligence, pages 197--202, Vancouver, B.C., 1981. [5] Velásquez, J. Modeling Emotions and Other Motivations in Synthetic Agents. In: Proceedings of AAAI-97, 1997 http://citeseer.ist.psu.edu/103027.html [6] Minar, N., R. Burkhart, C. Langton, and M. Askenazi. The Swarm simulation system: A toolkit for building multi-agent simulations. Working Paper 96-06-042, Santa Fe Institute, 1996 [7] D.A. Holway, L. Lach, A.V. Suarez, N.D. Tsutsui, and T.J. Case. The causes and consequences of ant invasions. Annual Review of Ecological Systems. 33:181-233, 2002. [8] D. A. Holway, A.V. Suarez, and T.J. Case. Loss of intraspecific aggression in the success of a widespread invasive social insect. Science, vol. 282:30, 949-952, 1998. [9] D.A. Holway. Competitive mechanisms underlying the displacement of native ants by the invasive Argentine ant. Ecology, 80:1, 238-351, 1999. [10] K.G. Human and D.M. Gordon. Exploitation and interference competition between the invasive Argentine ant, Linepithema humile, and native ant species. Oecologia, 105:405-412, 1996. [11] D.J.T. Sumpter and M. Beekman. From nonlinearity to optimality: pheromone trail foraging by ants. Animal Behaviour, 66:273-280, 2003. [12] M. Scheutz. The evolution of simple affective states in multi-agent environments. In; Proceedings of AAAI-01, 2001. [13] D.W. Davidson. Resource discovery versus resource domination in ants: a functional mechanism for breaking the trade-off. Ecological Entomology, 23:484-490, 1998. [14] D.R. Chialvo and M.M. Millonas. How swarms build cognitive maps. Biology and Technology of Intelligent Autonomous Agents, 144:439-450, 1995. [15] Matthias Scheutz and Paul Schermerhorn. The more radical, the better: Investigating the utility of aggression in the competition among different agent kinds. In Proc. of SAB 2004, Los Angeles, CA. MIT Press, 2004.