Social learning in a simple task allocation game

02/19/2017
by   Rui Chen, et al.
Monash University
0

We investigate the effects of social interactions in task al- location using Evolutionary Game Theory (EGT). We propose a simple task-allocation game and study how different learning mechanisms can give rise to specialised and non- specialised colonies under different ecological conditions. By combining agent-based simulations and adaptive dynamics we show that social learning can result in colonies of generalists or specialists, depending on ecological parameters. Agent-based simulations further show that learning dynamics play a crucial role in task allocation. In particular, introspective individual learning readily favours the emergence of specialists, while a process resembling task recruitment favours the emergence of generalists.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 2

09/04/2021

A Case Study of Agent-Based Models for Evolutionary Game Theory

Evolutionary game theory is a mathematical toolkit to analyse the intera...
04/08/2019

Coupling agent based simulation with dynamic networks analysis to study the emergence of mutual knowledge as a percolation phenomenon

The emergence of mutual knowledge is a major cognitive mechanism for the...
06/04/2020

How individual behaviors drive inequality in online community sizes: an agent-based simulation

Why are online community sizes so extremely unequal? Most answers to thi...
02/01/2018

Evolutionary Model Discovery of Factors for Farm Selection by the Artificial Anasazi

Agent-based modeling has been criticized for its apparent lack of establ...
04/25/2017

The Emergence of Consensus: A Primer

The origin of population-scale coordination has puzzled philosophers and...
10/08/2001

Alife Model of Evolutionary Emergence of Purposeful Adaptive Behavior

The process of evolutionary emergence of purposeful adaptive behavior is...
07/10/2018

Dynamics of Taxi-like Logistics Systems: Theory and Microscopic Simulations

In this paper we study the dynamics of a class of bi-agent logistics sys...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Background

The ability of social insect colonies to tackle different tasks simultaneously without central control is a key factor in their ecological success. Individual choices by colony members give rise to collective behaviours that are robust to dynamic and uncertain environments (Camazine et al., 2001). The mechanisms of this process, known as self-organised task allocation, have inspired a variety of applications in Computer Science, including scheduling (Campos et al., 2000) and control (Dressler, 2007).

Individual choices in task allocation are determined by three different interacting factors: individual personalities, the environment, and interactions amongst colony members (Charbonneau and Dornhaus, 2015). The interplay between individual traits and the environment has been studied extensively using models (Bonabeau et al., 1996; Duarte et al., 2012; Lichocki et al., 2012); but the effect of social interactions is less well-understood  (Jeanson and Weidenmüller, 2014; Kao et al., 2014).

We investigate the effects of social interactions in task allocation using Evolutionary Game Theory (EGT). We propose a simple task-allocation game and study how different learning mechanisms can give rise to specialised and non-specialised colonies under different ecological conditions. By combining agent-based simulations and adaptive dynamics we show that social learning can result in colonies of generalists or specialists, depending on ecological parameters. Agent-based simulations further show that learning dynamics play a crucial role in task allocation. In particular, introspective individual learning readily favours the emergence of specialists, while a process resembling task recruitment favours the emergence of generalists.

2 A Simple Task-Allocation Game

Our model assumes that individuals live in groups of size , and choose to allocate their effort into two different tasks: Task A is a regulatory task, such as fanning to cool down nest temperature; Task B is a foraging task, required to support the energy costs of the colony. The behaviour of individual is determined by a continuous trait

, the probability that individual

will perform Task A. Thus, is the probability of individual engaging in Task B. In the most general sense, the payoff of individual is given by: , where and are benefits and cost respectively. Total payoff for one individual depends not only on her own trait, but also on the traits of all others in the population.

We define the benefit function as . This implies that the benefits of foraging are discounted if the colony is not well-regulated. For the regulatory task, we assume that benefits are maximised when an intermediate number of individuals are engaged in regulation, thus is a concave function in . For the foraging task we assume that individual benefits increase linearly with collective foraging effort.

To fix ideas is assumed to be quadratic with maximum value 1 when half of the workers in the group are engaged in Task A. Likewise, is the individual reward from performing Task B in one time period. Thus,

and

For this payoff structure the rewards of foraging are maintained in full, only when regulation is optimal.

Costs are defined as . The cost of regulation is assumed to be linear, with . Here, represents the cost of performing Task A for an individual in one time period. The nature of foraging implies decreasing marginal costs, thus for .

3 Social learning

How do individuals in the colony learn to coordinate their efforts to successfully perform both tasks? We obtain a first approximation by using the mathematical framework of adaptive dynamics (Brännström et al., 2013). The underlying process resembles social learning, whereby individuals tend to copy those that are most successful. The assumptions of adaptive dynamics imply a very large population where mutations are rare and small (Waxman and Gavrilets, 2005).

The analysis predicts a unique singular strategy that is convergent stable. We derive conditions for evolutionary branching; a situation in which a monomorphous population splits in two morphs (Doebeli et al., 2004). In the context of task allocation, this implies that the colony successfully tends to both tasks, but each individual specializes by putting all their effort into one single task. Alternatively, the colony can converge to a state in which generalists share responsibilities and each individual splits efforts in both tasks. We refer to the latter as weak specialisation, as opposed to strong specialisation in the former case. Given , the colony exhibits strong specialisation if

or weak specialisation (evolutionary stable strategy) if

The system can also converge to a state in which the colony fails to coordinate its efforts in both tasks. In this case we speak of an inviable colony.

Figure 1 shows how these different outcomes depend on ecological conditions as given by the values of and .

Figure 1: Classification of colony behaviour as predicted by adaptive dynamics
Figure 2: Agent-based simulations for social learning

To check the robustness of our mathematical prediction we use an agent-based simulation. Social learning follows a Wright-Fisher process where the population is not necessarily monomorphous (Imhof and Nowak, 2006)

. This is similar to roulette wheel selection in a standard Genetic Algorithm

(Fogel, 2006).

Figure 2 shows that our simulation results are in line with the predictions from adaptive dynamics. Our results indicate that when foraging is expensive, agents fail to coordinate and the colony is not viable. For intermediate costs of foraging individuals strongly specialize, while cheap foraging leads to a population of generalists.

4 Other Learning Mechanisms

The mathematical framework and the simulations so far assume that agents learn socially. This implies that individuals can observe the traits of all others in the colony, while at the same time inferring how successful a particular trait is. We perform simulations in which this assumption is relaxed in two different ways: i) A model of individual learning assumes that agents learn exclusively from their own experience. ii) A model of task recruitment assumes that agents respond to recruitment signals for either task.

4.1 Individual Learning

For individual learning we assume that agents remember their most recent strategy and its associated payoff. Innovations arise via exploration with a small probability. If a new strategy is not better than the most recent one, individuals roll back to their previous strategy. This class of introspective learning is thought to be less cognitively demanding than social learning.

Figure 3: Classification of colonies based on the model of individual reinforcement

Figure 3 shows the simulation results for this model. Individual learning readily favours strong specialisation across a large range of parameters.

4.2 Task Recruitment

We also investigate a process of task recruitment, inspired by the ecological literature on task allocation (Dukas, 2008). In this version of the model, agents respond to recruitment signals for each task. At each time-step, individuals choose a single task to perform based on their trait, i.e., individual chooses Task A with probability . In a recruitment phase, individuals modify their trait according recruitment signals send by successful individuals in the population. In particular, we assume that the intensity of the signal is proportional to the fitness of the recruiter.

Figure 4: Classification of colonies based on the model of task recruitment

The simulation results for this model are shown in Figure 4. Task recruitment favours weak specialisation across a large range of parameters.

5 Discussion

This paper shows that EGT is a promising avenue to study the effect of social interactions in task allocation models. In particular, we show that a simple model of social learning can give insights into when to expect different levels of specialization. We show that the mechanism by which individuals learn can dramatically change predictions: while social learning is conducive to societies of specialists and generalists, individual learning readily leads to strong specialization.

More research is needed to understand how different species may rely on different kinds of learning. Our models suggest that the ecology of the tasks interacts in a non-trivial way with the cognitive capacities of the species.

Extensions of this model can introduce different types of tasks as well as the option of performing no tasks, which has been widely observed in social insect colonies. Further research is also needed to provide analytic predictions for individual learning, and to understand how different time-scales of learning and ecology may interact with each other in task allocation games.

References

  • Bonabeau et al. (1996) Bonabeau, E., Theraulaz, G., and Deneubourg, J.-L. (1996). Quantitative study of the fixed threshold model for the regulation of division of labour in insect societies. Proc. R. Soc. Lond. B.
  • Brännström et al. (2013) Brännström, Å., Johansson, J., and von Festenberg, N. (2013). The Hitchhiker’s Guide to Adaptive Dynamics. Games 2013, Vol. 4, Pages 304-328.
  • Camazine et al. (2001) Camazine, S., Deneubourg, J.-L., Franks, N. R., Sneyd, J., Theraulaz, G., and Bonabeau, E. (2001). Self-organization in biological systems. Princeton University Press.
  • Campos et al. (2000) Campos, M., Bonabeau, E., Theraulaz, G., and Deneubourg, J.-L. (2000). Dynamic scheduling and division of labor in social insects. Adpt. Behav.
  • Charbonneau and Dornhaus (2015) Charbonneau, D. and Dornhaus, A. (2015). When doing nothing is something. How task allocation strategies compromise between flexibility, efficiency, and inactive agents. J Bioecon.
  • Doebeli et al. (2004) Doebeli, M., Hauert, C., and Killingback, T. (2004). The evolutionary origin of cooperators and defectors. Science, 306(5697):859–862.
  • Dressler (2007) Dressler, F. (2007). Self-organization in sensor and actor networks. Wiley, Chichester, England.
  • Duarte et al. (2012) Duarte, A., Pen, I., Keller, L., and Weissing, F. J. (2012). Evolution of self-organized division of labor in a response threshold model. Behav. Ecol. Sociobiol.
  • Dukas (2008) Dukas, R. (2008). Evolutionary Biology of Insect Learning. Annu. Rev. Entomol.
  • Fogel (2006) Fogel, D. B. (2006). Evolutionary computation: toward a new philosophy of machine intelligence, volume 1. John Wiley & Sons.
  • Imhof and Nowak (2006) Imhof, L. A. and Nowak, M. A. (2006). Evolutionary game dynamics in a wright-fisher process. Journal of mathematical biology, 52(5):667–681.
  • Jeanson and Weidenmüller (2014) Jeanson, R. and Weidenmüller, A. (2014). Interindividual variability in social insects - proximate causes and ultimate consequences. Biol. Rev. Camb. Philos. Soc.
  • Kao et al. (2014) Kao, A. B., Miller, N., Torney, C., Hartnett, A., and Couzin, I. D. (2014). Collective learning and optimal consensus decisions in social animal groups. PLoS Comput. Biol., 10(8):e1003762.
  • Lichocki et al. (2012) Lichocki, P., Tarapore, D., Keller, L., and Floreano, D. (2012). Neural networks as mechanisms to regulate division of labor. Am. Nat.
  • Waxman and Gavrilets (2005) Waxman, D. and Gavrilets, S. (2005). 20 questions on adaptive dynamics. Journal of evolutionary biology, 18(5):1139–1154.