Drivers for the Evolution of Moral. Behaviour in Multi-Agent Societies

(1)

Drivers for the Evolution of Moral

.

Behaviour in Multi-Agent Societies

Jeroen Rijnbout 10456341

Bachelor thesis Credits: 18 EC

Bachelor’s Programme Artificial Intelligence

University of Amsterdam Faculty of Science Science Park 904 1098 XH Amsterdam Supervisor dhr. dr. B. Bredeweg Informatics Institute Faculty of Science University of Amsterdam June 26th, 2015

(2)

Abstract

This paper investigates the effects of various drivers on the evolution of moral behaviour to determine whether the theory of evolutionary ethics can be considered a valid explanation for the existence of moral behaviour. A canonical genetic algorithm is used to simulate evolution in a multi-agent system. Several drivers are studied to determine under what conditions soci-eties can evolve moral behaviour. The models show that when co-operation is necessary to fully exploit the environment’s resources, societies evolve more co-operative behaviour. In addition, results show that family relations combined with individual illness drive societies to evolve altruistic behaviour.

(3)

1 Introduction

1.1 Theoretical Background

The theory of evolutionary ethics posits that the human moral sense emerged through natural selection. If this were true, morality could be understood as a phenomenon that arises automatically during the evolution of sociable, intelligent beings and not as the result of the application of our rational fac-ulties. Evolutionary ethics interprets morality as a favourable adaptation to the environment which increases the fitness of individuals that base their actions upon good morals. Therefore, this selective advantage means moral individuals have a higher survival rate and are more likely to be selected in the evolutionary cycle (Schroeder, 2015). As a result, moral behaviour spreads to future generations and becomes a dominant trait within a popu-lation.

The idea of moral principles as a genetically built-in mechanism origi-nated from the works of Charles Darwin. In his work The Descent of Man, Darwin (1871) argued that human morality stems from the social instincts that developed in some lifeforms during evolution. From a biological point of view, this claim can be explained as follows. Early lifeforms, such as amoebae and frogs, did not require sociability to ensure the survival of their offspring. The amoebae reproduce by division and frogs abandon their tadpole-offspring, leaving them to fight for their own survival. However, the emergence of more advanced lifeforms, such as birds, gave rise to the parental instinct; birds brood their eggs and selflessly feed their young to ensure survival of their offspring. These parental responsibilities required social mechanisms unseen in earlier evolutionary history (Schroeder, 2015). It is here that evolutionary ethics steps in and tries to explain the emergence of moral behaviour from an evolutionary approach.

Modern computing techniques enable researchers to test evolutionary theories by using genetic algorithms and multi-agent systems to simulate evolution. These tools can be used to approach the theory of evolutionary ethics from a computational point of view and help determine whether it is a valid explanation for the existence of moral behaviour. Previous work using this approach has shown that egoistic agents tend to perform poorly in long term simulations, even being unable to survive and going extinct (Bazzan, Bordini, & Campbell, 2002). In addition, altruistic agents were found to be able to have a positive impact on the social group they were in while only minimally compromising their individual performance (Bazzan, Bordini, & Campbell, 1999). Furthermore, moral or altruistic behaviour can indeed evolve in simple multi-agent societies given the right circumstances. These results indicate that the theory of evolutionary ethics may indeed be a valid explanation for the human moral sense. However, not all societies automatically evolve a moral sense. In testing family relations, individual

(5)

memory and agent reputation as drivers for the evolution of a moral sense, Spronck and Berendsen (2009) found that only specific configurations of societies with family relations or agent reputation evolved a moral sense and general models did not. Therefore, it was concluded that evolving moral behaviour may not be as obvious as it seems.

1.2 Drivers

Simulating the effects of additional drivers, such as a constraint on food gathering, resource scarcity (famine) and individual illness, could help to explain the natural selection of moral behaviour and help support the theory of evolutionary ethics.

1.2.1 Food Constraint

Co-operating to achieve a common goal can often lead to much greater per-sonal benefit than competing for limited resources (Ruse, 1986). This gives rise to the question whether humans evolved to be collaborative species due to a need for teamwork in gathering food. Hunter-gatherer societies, which sustained humanity during the palaeolithic era up to the advent of agricul-ture some 10 000 years ago, relied heavily on animal foods (i.e. meat). Most hunter-gather societies derived over 50% of their subsistence from animal foods (Cordain et al., 2000). However, while some foods can be gathered and consumed by an individual, such as plant foods and fished animal foods, many require coordinated efforts, such as big game hunting. Therefore, it is relevant to investigate whether co-operation as a requirement for food gathering drives a society towards a more altruistic and moral inclination. 1.2.2 Famine

Besides co-operating to ensure being able to gather sufficient sustenance, it is possible that the act of sharing with other group members affected survival. Surviving periods of resource scarcity or famine by sharing what is available could have been a driver for the evolution of morality, as recurrent scarcity was a fact of life for hunter-gatherer societies (Colson, 1979). Furthermore, hunter-gatherer societies were heavily dependent on stored foods; periods of abundance were separated by periods of scarcity where stored foods served as the main source of sustenance (Keeley, 1988; Smith, 2003). Perhaps in-dividuals in hunter-gatherer societies altruistically shared their food with other members in their society to survive these periods of resource scarcity as a group. However, the sharing of meat with other group members has not been sufficiently accounted for by previous models (Henrich, 2012). There-fore, the effects of altruistic sharing on group survival will need to be studied to determine whether resource scarcity can be regarded as a valid driver for the evolution of altruism.

(6)

1.2.3 Individual Illness

Another factor that may have affected evolution towards altruistic behaviour is the inability to gather food for oneself due to an individual disability, such as physical trauma (e.g. a broken leg) or infectious disease. As hunter-gatherer societies lived in relatively small groups, they were less prone to suffer from infectious diseases that require large and dense groups and more likely to be impacted by individual trauma (Hill, Hurtado, & Walker, 2007). When a person is incapacitated due to individual illness or injury, this in-dividual will need to be cared and provided for during recovery. Likewise, to prevent diminishing group numbers, that individual will need to care for other members of the group when they are somehow unable to provide for themselves. This mutual care-taking is beneficial for individual survival and, consequently, also for group survival. It is likely that this mechanism is based on altruistic reciprocity, where an individual, A, is more likely to be altruistic towards another individual, B, if B has behaved altruistically to-wards A in the past (Falk & Fischbacher, 2006; Trivers, 1971). In addition, saving an individual’s life could enable co-operative hunting to continue with more individuals (Trivers, 1971). Therefore, individuals have an incentive to care for and heal other members so they can continue hunting in larger groups and enjoy larger gains. To determine whether individual illness is in fact a valid driver for the evolution of altruism, the effects of altruistic reciprocity and individual illness on group survival should be studied.

2 Simulating Evolution

To study the effects of the drivers described in section 1.2, several models were built that simulate evolution. The models simulate hunter-gatherer societies in a multi-agent system using a genetic algorithm. This section describes how the models were implemented, how the environment and so-cieties within the models are represented and how the genetic algorithm works. The implementation was done in NetLogo, a programmable multi-agent modelling environment for simulating natural and social phenomena (Wilensky, 1999). Five models were built to study the effects of each driver: food constraint (2.5.1), famine (2.5.2), individual illness (2.5.3) and nuclear family relations combined with famine and individual illness (2.5.4).

2.1 Basic Environment

The basic design for each model consists of an environment in which the agents operate, the society (i.e. a group of agents) and the actions they can perform. The environment is set on an 11 × 11 grid through which agents can move freely (the environment wraps around the edges). Each patch represents an area with or without food; food is randomly distributed

(7)

among all patches upon initialization of the model, with each patch having a 20% chance of being assigned food. See Figure 1 for a visual representation of the environment; blue patches represent areas with food and brown patches represent empty areas. When a patch of food is collected by an agent, the patch is emptied and turns brown. Every iteration each brown patch then has a chance to receive new food, this chance varies per model as environments with some drivers require more resources for basic survival.

Figure 1: Model environment, 11 × 11 grid.

2.2 Agent Attributes and Actions

The environment contains a society of agents that all have their own at-tributes. Some of these attributes are the same for each model:

• an agent’s energy (i.e. an agent’s health) • an agent’s age

• the probability an agent will select an egoistic action

• the probability an agent will select an altruistic/co-operative action Agents are able to select two types of actions: an egoistic action, where agents generally only collect food for themselves, and an altruistic or co-operative action, where agents share their resources or work together to gather food. The probabilities for either type of behaviour are represented by two decimal numbers between 0 and 1 (e.g. 0.35 and 0.65). Agents select actions by randomly generating a number between 0 and 1. If the num-ber is equal to or below the chance that they will select an egoistic action, an egoistic action is selected. If the generated number is above the egois-tic probability, an altruisegois-tic action is chosen. In addition, agents also have model-specific personal attributes; these are detailed in following sections.

(8)

Upon initializing the model, agents are randomly placed across the envi-ronment. Patches do not correlate to agent placement. Therefore, multiple agents can occupy the same patch. Figure 2 depicts a fully initialized model.

Figure 2: Model environment, 11 × 11 grid, including agents.

Each iteration an agent will start by moving one step forward in a ran-dom direction, to enable agents to explore the world and gather food. This step consumes one unit of energy of an agent’s total energy. Next, agents will check whether the patch they are on contains food and, if food is present, gather it. This increases an agent’s energy by ten. The next step is re-production, where agents find a suitable mate and produce offspring which carries a combination of its parents’ genes. See section 2.4 for a detailed description of the reproduction process. After reproduction, every agents’ age increase by one as they grow older. Finally, agents whose health is below zero die and are removed from the environment. In addition, agents that have grown older than the maximum age (>60) die as well.

2.3 Hunter-Gatherer Societies

To model the conditions under which moral behaviour is likely to have evolved, model populations are based on hunter-gatherer societies. The hunter-gatherer way of life serves as a simplistic abstraction of a society without the complexities of modern life. Therefore, it is suited for use in an evolutionary simulation. To correctly recreate hunter-gatherer societies, some parameters have been based on what we know of their lifestyle.

Firstly, the environment size (11 × 11 grid) and the introduction of new resources have been set so that the environment supports societies with an average group size of around 80 members. This group size is likely to approx-imate the number of members hunter-gather societies during the palaeolithic era (Choi & Bowles, 2007). However, societies are initialized with a group

(9)

size of 30 members; this is done so the group can grow to the environment’s maximum carrying capacity if the society performs well. Secondly, if children survived past the age of 15, the average age of death for hunter-gatherers was around 60 years old (Kaplan, Hill, Lancaster, & Hurtado, 2000; Gurven & Kaplan, 2007). Therefore, the maximum age agents can grow to has been set to 60 years old. Finally, we consider reproduction ages and reproduction frequency. Research has shown that the average ages of first and last birth in hunter-gatherer societies are around 18 and 34 years old, respectively (Fenner, 2005). Therefore, agents in the model can only start reproducing when they reach 18 years of age and stop reproducing when they reach the age of 34. In addition, hunter-gatherer women have children spaced about four years apart to reduce the load women have to carry while the group is foraging (Jones, 1986). Accordingly, agents in the model are only eligible for reproduction once every four years.

2.4 Genetic Algorithm

Reproduction is based on the canonical genetic algorithm described by Whitley (1994), which consists of three main steps: selection (2.4.1), re-combination (2.4.2) and mutation (2.4.3). This algorithm is represented and detailed below. It starts with the current population. Selection is then applied to the current population to create the intermediate population. The next population is then generated by applying recombination and mutation to the intermediate population. An overview of the algorithm is shown in Algorithm 1.

2.4.1 Selection

The algorithm starts by determining whether an agent is fit for reproduction. Selection criteria are based on hunter-gatherer reproduction habits detailed in section 2.3 and are as follows. Firstly, an agent has to have at least 50 units of energy to reproduce. This represents the agent’s fitness and a certain level is required to be able to survive reproduction. When an agent reproduces, their health is reduced by half as reproduction requires a considerable investment of energy. Therefore, an agent’s energy has to be above 50 units of energy for it to able to survive reproduction; reproducing at lower energy levels could leave the agent unable to survive the next few iterations. Secondly, an agent has to be at least 18 years of age to be considered sexually mature and ready for reproduction. Thirdly, an agent cannot be over 34 years of age as this is close to the average age of last birth for hunter-gatherers (section 2.3). Finally, an agent cannot have reproduced in the past four years to be eligible for reproduction. If an agent that is looking to reproduce is deemed fit and eligible for reproduction it becomes the first parent (P 1). A suitable partner is then determined by randomly

(10)

selecting another agent from the society that meets the selection criteria, this becomes the second parent (P 2). Note that individuals that are very fit, i.e. they are good at collecting resources, will likely reproduce more; if an agent is eligible for reproduction but does not have enough resources it will not reproduce.

Algorithm 1 Genetic algorithm procedure Reproduction

procedure selection ← current population

select two parents that meet reproduction requirements: energy >= 50

age >= 18 age <= 34 eligible = true

procedure recombination ← parents’ genes

combine parent 1 (P 1) and parent 2 (P 2) genes into child genes: child genes = (P1 + P2) / 2

procedure mutation ← child genes if random int 100 < 3 then

mutate child genes by 5%

2.4.2 Recombination

When two agents, P 1 and P 2, have been selected for reproduction, recombi-nation of their genes takes place. As described in section 2.2, an agent’s genes are represented by its chance for exhibiting either egoistic or co-operative and altruistic behaviour. Therefore, the recombination of the genes carried by P 1 and P 2 becomes the average of their respective chances for exhibiting a certain type of behaviour. For example, as depicted in Figure 3, if P 1 has an egoistic chance of 0.7 and P 2 has an egoistic chance of 0.1, the child egoistic chance becomes 0.4 through recombination.

Figure 3: Recombination of genes (before possible mutation). Red repre-sents egoistic chance, green reprerepre-sents altruistic chance.

(11)

2.4.3 Mutation

The final step in the reproduction process is that of mutation. Mutation serves to maintain genetic diversity within the society, which prevents the algorithm from getting stuck in local minima. This also enables search-optimization in searching for the fittest genes and prevents the society’s gene-pool from becoming too similar, which stops evolution. The mutation rate is set to 3%, i.e. in 3% of all children the recombined genes are slightly altered. When genes are mutated, they are altered by 5% and there is a 50% that either the egoistic or altruistic gene will be increased. Therefore, if a child’s genes are mutated, an egoistic chance of 0.7 has a 50% chance of being increased by 5% through mutation to 0.75. Consequently, the corresponding altruistic chance will decrease by 5%. After mutation a child’s genes are final and the child is put into the society. Agents can evolve to become either fully egoistic: a 1.0 probability for exhibiting egoistic behaviour, or fully altruistic: a 1.0 probability for exhibiting altruistic behaviour.

2.5 Models

This section discusses the different types of models that were built to test the effects of the various drivers on what behaviours a society evolved. Firstly, section 2.5.1 discusses the implementation of a model where two types of resources are required for survival. Secondly, section 2.5.2 discusses the im-plementation of a model where periods of abundance alternate with periods of famine. Thirdly, section 2.5.3 discusses the implementation of a model where individual illness is present among members of the society. Finally, section 2.5.4 discusses the addition of nuclear family relations to the models of famine and individual illness.

2.5.1 Food Constraint

The mechanism that drives this model is a distinction between two types of food: one that is solo-gatherable and one that is co-op-gatherable. Whereas the other models only have one type of food, which can be gathered by indi-vidual agents, this model requires agents to work together to gather a second food source which can only be gathered through a collaborative effort. There are some resources that can be gathered by an individual agent, but these are insufficient for survival. Therefore, both types of food need to be collected to survive. Each type of food has a 10% chance of being assigned to a patch. It is hypothesized that a constraint on gathering different types of food will cause agents to evolve towards solely performing co-operative actions, as this increases their chances of gathering resources and, consequently, of re-production. Figure 4 illustrates the addition of the second type of resource: a resource that can only be gathered by co-operation, shown in pink.

(12)

Figure 4: Two types of resources: solo-gatherable (blue) and co-op-gatherable (pink).

In this model, when an agent selects an egoistic action the agent will simply check if there is a resource available on its current patch which it can collect by itself and, if so, collect it. However, if an agent selects a co-operative action the agent will perform an additional step. Like the egoistic action, it will check whether there is a resource available that the agent can gather by itself. In addition, if the agent finds there is a resource on its current patch that can only be gathered through co-operation the agent will check the eight adjacent patches to see if there are any agents present. If one or more agents are present on the surrounding patches they will also select an action, if the action they select is co-operative they will help the agent at play collect the food. The agents that helped the agent at play collect the food are then incurred a cost for the energy they expended in the process, which represents the altruistic sacrifice an individual makes to help another group member. This cost is the amount of energy the agent at play gained, divided by the number of agents that helped; this is the cost of gathering a co-operative resource.

2.5.2 Famine

In this model every agent has an equal chance of gathering food; there is only one type of solo-gatherable food. However, periods of famine are introduced. The mechanism that drives this model is periods of resource scarcity. There is a 15% chance a patch is assigned new food, as a lower setting results in extinct societies. Resources are abundant for 25 iterations, followed by 5 iterations where no new resources are introduced into the model. This process is then repeated for the duration of the model. Agents will need to survive periods without the introduction of new resources into

(13)

the environment. Figure 5 shows the amount of resources available during the first few iterations of a run. Periods of scarcity are clearly visible every 25 iterations, where the amount of total resources in the environment dips (close) to 0.

Figure 5: Periods of resource scarcity are introduced every 25 iterations. Agent behaviour types are defined as: egoistic and altruistic. If an agent selects an egoistic action, it will attempt to gather resources from the patch it is currently on. However, when an agent selects an altruistic action the agent will collect food from its patch if possible and then share half of the gathered resources with the agent that is lowest on energy of all agents. Although, the agent will only share its resources if this does not bring the agent’s energy dangerously low (<10 units of energy). This sharing behaviour represents the altruistic sacrifice of sharing resources with other less-fortunate agents. It is hypothesized that agents will need to share with each other to be able to survive the periods without resources. Therefore, it is expected that agents will develop towards selecting predominantly sharing behaviour.

2.5.3 Individual Illness

As in the model described in section 2.5.2, every agent in this model has an equal chance of gathering the solo-gatherable resources that are available. However, the mechanism that drives this model is individual illness. Periods where an agent is incapable of collecting food for itself are introduced. Every iteration an agent is randomly selected to become ill and incapacitated for 10 iterations, rendering them unable to gather food for themselves while still consuming resources. As resources are constantly available, less are required to avoid total extinction. Therefore, each patch has a 10% chance of being assigned new food. The idea is that sick agents will need other agents to share their resources with them to ensure their survival. Figure 6 depicts an environment with several sick agents in it; they are unable to move until they are cured (i.e. survived 10 iterations).

(14)

Figure 6: Sick agents are represented by their red color and the inability to move.

Here, healthy agents are able to share some of their food with sick agents. If an agent has selected an altruistic action, the agent will collect the avail-able resources on its patch and determine whether it is healthy enough to share some of its food. Agents are considered healthy and eligible for shar-ing if they are able to share and still maintain an energy level fit enough for reproduction (i.e. >50). Therefore, if an agent has more than 60 units of energy it will share half of the resources it gathered that round into a global food pot. The resources in this pot can then be accessed by sick agents that have selected an altruistic action, even though they are unable to gather food they will now gain some energy by taking resources from the food pot. This mechanism is based on reciprocity theory, where agents that share a lot (and thus make many altruistic sacrifices) are also more likely to be shared with if they are in need. It is hypothesized that agents will evolve sharing behaviour, as this is what enables them to survive periods of individual illness.

2.5.4 Nuclear Family Relations

In addition to the previously described drivers, family relations were added to the models described in sections 2.5.2 and 2.5.3, as Spronck and Berend-sen (2009) found that family relations had an impact on what behaviours societies evolved. Now, when a child is born, links are created between the child and its parents. This enables parents to care for their children by sharing some of the resources they gather each round with their children. Agents that select an egoistic action now distribute half of the resources they gather each round among their children, while still not sharing any resources with the rest of society. Likewise, agents that select an altruistic action now also distribute half of their resources among their children. In

(15)

addition, these agents also donate one quarter of the resources they have gathered each round by either 1) sharing with the agent lowest on energy, in the environment where famine is present or 2) by donating to the food pot in the environment where individual illness is present. Figure 7 shows a model in which family relations have formed after several iterations.

Figure 7: Nuclear family relations: the directed arrows represent a parent-to-child relation.

It is hypothesized that family relations will provide egoists with the same advantage sharing agents have, without the disadvantage of sharing with non-related members of society. Therefore, it is expected that egois-tic societies will become stronger and societies that previously evolved an altruistic sense are now less likely to do so.

3 Results

This section discusses the results of the simulations and how they were evaluated. Section 3.1 details how results were gathered to evaluate the models. Sections 3.2, 3.3, 3.4 and 3.5 discuss the results of their respective models.

3.1 Evaluation

To evaluate whether societies evolved an egoistic or an altruistic sense, sim-ulations were run 100 times per model. Models were run for either 50 000 or 100 000 iterations per run, depending on how long they took to con-verge towards a specific behaviour. Results from these 100 runs are then used to determine mean values for what sense the societies evolved and conclude whether a driver causes societies to evolve a moral sense. These values can also indicate what type of behaviour is optimal for survival

(16)

un-der the varying conditions of each model and give an indication of whether mankind evolved co-operative and altruistic behaviour because it increased our chances for survival.

3.2 Food Constraint

The co-operative model was run for 100 000 iterations per run, as this was re-quired to enable societies in the model to evolve towards optimal behaviour. Out of 100 runs, 33 societies went extinct due to evolving an egoistic sense after an average of 23 748 iterations (co-operation was a requirement for survival). Table 1 shows the results for all 100 runs of this model, which include the last measured values for extinct societies. These results show a clear tendency towards evolving co-operative behaviour, with an average co-operative action selection chance of 0.65.

Average St. dev. Min. Max. Co-operative 0.65 0.2 0.35 0.97

Egoistic 0.35 0.2 0.03 0.65 Table 1: Results for 100 of runs of the co-operative model.

However, if we only consider the runs that survived the full 100 000 iterations and do not take into account the societies that went extinct the results show societies evolved decidedly co-operative behaviour. Table 2 shows the results for the societies that managed to survive the entire run.

Average St. dev. Min. Max. Co-operative 0.77 0.13 0.52 0.97

Egoistic 0.23 0.13 0.03 0.48 Table 2: Results for the 67% of runs that did not go extinct.

These results suggest that a need to work together in collecting sufficient sustenance could indeed be a driver for the emergence of moral behaviour. However, agents did not evolve to become fully co-operative; they still have an egoistic selection chance of 0.23 on average. This could be due to some models requiring even more runs to converge to the optimal behaviour or it could be that some amount of egoism is beneficial to individual agents. By having some chance for gathering resources that can be gathered by an individual agent, societies make full use of all resources available - whereas many resources would be wasted if they evolved to solely collect co-op-gatherable food.

(17)

3.3 Famine

The model with recurrent periods of scarcity was run for 50 000 iterations per run, as this was sufficient time for societies to evolve a distinct prefer-ence for one type of behaviour. This reduction in the number of iterations required would suggest that resource scarcity acts as a stronger, more dras-tic driver for the selection of optimal behaviour. In this case, as can be seen in Table 3, societies evolved decidedly egoistic behaviour; an average chance for selecting egoistic behaviour of 0.93 and a low standard deviation indicate that behaving egoistically is the optimal survival/reproduction strategy for this model.

Average St. dev. Min. Max. Altruistic 0.07 0.09 0.02 0.53

Egoistic 0.93 0.09 0.47 0.98

Table 3: Results for 100 runs of 50 000 iterations where periods of famine are present.

Surprisingly, there was no mass starvation during each period of scarcity; no agents died of starvation. This is likely due to agents sharing with the agent that is lowest on health, thereby saving it from starvation. It appears that a shortage in food mainly affects an agents’ chances of reproduction. An altruistic agent that shares more will not be able to reproduce after hav-ing exhausted its resources in survivhav-ing without new food income, whereas an egoistic agent that hoarded all its food is able to reproduce. In addi-tion, it seems that egoists thrive due to leeching off of the resources shared by altruists. These results indicate that famine is a strong driver for the evolving egoistic behaviour within societies. It is likely that this is due to there being no direct benefit from sharing one’s resources with other agents; egoistic behaviour is beneficial as there are no repercussions.

3.4 Individual Illness

Individual illness was simulated for a duration of 100 000 iterations per run. One agent was made sick each iteration, resulting in a total of 100 000 sick agents during the simulation. Out of this total amount of sick agents, an average of 29,8% was cured (i.e. survived their illness period) through being cared for by others. This low percentage of cured group members correlates to the societies mainly evolving an egoistic sense, with an average chance for selecting an egoistic action of 0.67 (Table 4).

Contrary to what was hypothesized, agents in this model did not evolve a predominantly altruistic sense. These results indicate that individual illness alone cannot be considered as a factor that drives the evolution of altruistic

(18)

Egoistic 0.67 0.22 0.06 0.96 Table 4: Evolved behaviours driven by individual illness.

behaviour. However, variance is high (0.22 standard deviation) and the min-imum and maxmin-imum values for evolved behaviours are far apart. Therefore, these results are inconclusive in determining what effect individual illness has on the evolution of altruistic behaviour. Unexpectedly, starvation was present; an average of 35,9% of all agents starved to death. This is likely due to selectively sharing resources with those that have been kind before (i.e. reciprocity) instead of sharing with the agent with lowest health (as in the famine model); this allows agents that rarely select an altruistic action to die of starvation as they also have low chances of being shared with. This suggests that reciprocity does indeed increase an individual’s chances for survival and diminishes it for those that do not show similar kindness.

3.5 Nuclear Family Relations

Simulations for models where family relations were added were run the same number of iterations as in the famine and individual illness models; famine in combination with family relations and individual illness in combination with family relations were simulated for 50 000 and 100 000 iterations, re-spectively.

3.5.1 Famine combined with Family Relations

As in the model for famine without family relations, societies developed a decidedly egoistic sense; an average egoistic action selection chance of 0.95 was evolved (Table 5). Therefore, it can be concluded that the addition of family relations does not have a positive effect on whether societies evolve an altruistic sense when influenced by periods of famine. If anything, so-cieties evolved to me more egoistic; although, only slightly, as the chance for selecting an egoistic action was only 0.02 higher than in the model with famine alone.

Egoistic 0.95 0.04 0.59 0.97

(19)

However, the addition of family relations does seem to affect population size. Societies grew larger as they evolved towards the optimal survival behaviour, even though the total available resources per iteration were not changed from the original famine model. Figure 8 illustrates this growth; evolution towards the most favourable behaviour is plotted against the total population size at each iteration.

Figure 8: Evolving towards the optimal survival behaviour correlates to larger societies in combination with resource scarcity (total population size₁₀₀ for scaling).

Note that there was no correlation between optimal behaviour and group size in models without the addition of family relations. This suggests family relations have a positive effect on the way groups manage the available re-sources. In addition, societies were able to survive longer periods of scarcity than societies without family relations. When the duration of each period without resources was gradually increased, societies were found to be able to consistently survive famines of up to 10 iterations; whereas longer du-ration’s resulted in regular extinction. Nevertheless, 10 iterations is double the period populations without family relations were able to survive. This further supports the notion that family relations have a favourable effect on societal resource management.

3.5.2 Invididual Illness combined with Family Relations

Results for individual illness combined with family relations are shown in Table 6. Recall that societies that had to deal with individual illness with-out family relations evolved a predisposition towards egoistic behaviour. In contrast, societies that had to deal with individual illness with the addition of family relations evolved a markedly altruistic sense. Societies evolved an average altruistic action selection chance of 0.95, with minimum variance

(20)

and minimum and maximum values close together.

Egoistic 0.05 0.01 0.03 0.1

Table 6: Results for individual illness in combination with nuclear family relations.

These results indicate that individual illness in combination with family relations is a clear driver for societies to evolve altruistic behaviour. Fur-thermore, societies in this model cured an average of 89.7% of their sick. This is a clear increase from the 29,8% from the model without added fam-ily relations. However, the percentage of agents that starved to death also increased: from 35,9% to 56%. Although, starvation did not affect popula-tion size. Together, these results suggest that societies with family relapopula-tions are able to more effectively cope with personal sickness but at a cost to the amount of starvation present in the society. It is likely that parents caring for their young make a personal sacrifice to keep them healthy, enabling the young to survive their illness and reproduce frequently to maintain popula-tion levels.

Figure 9: Evolving towards the optimal survival behaviour correlates to larger societies in combination with individual illness (total population size₁₀₀ for scaling).

Furthermore, similar to the presence of resource scarcity, family rela-tions appear to have a positive effect on group size. As shown in Figure 9, total population size increases as agents evolve towards the behaviour that makes them fittest for selection. Together with an increased number of cured

(21)

agents, this indicates that family relations do indeed affect the effectiveness of how societies handle the resources at their disposal.

4 Conclusion

The effects of various drivers on the emergence of moral behaviour within societies were studied by simulating evolution in a multi-agent system. A genetic algorithm was used to simulate reproduction and enable selection of the fittest behavioural traits. Results show that multiple factors can be considered valid drivers for the emergence of moral behaviour. Firstly, the requirement to collaborate in gathering two types of resources was found to have a positive effect on the level of co-operative behaviour societies evolved. Secondly, individual illness was found to favourably affect the level of altru-ism societies evolved when combined with family relations. Famine and individual illness alone did not cause societies to evolve a more altruistic in-clination; these societies evolved a decidedly egoistic tendency. In addition to the level of altruism societies evolved, it was found that family relations have a positive effect on the overall group size an environment can sup-port. Regardless of whether the evolved behaviour was egoistic or altruistic, population sizes increased as societies evolved towards the optimal survival behaviour. This is likely due to more effective resources management when family relations are present. Results support the theory of evolutionary ethics, as societies were able to evolve moral behaviour when pressured by various drivers. This contributes to the idea that the human moral sense could be the result of evolutionary ethics.

5 Discussion and Future Work

While this research has provided insights into what behaviour increases an individual’s chances for survival in specific conditions, the representation of behaviour selection is rather basic. Future research could distinguish more types of behaviour and allow for more complex interactions. For example, stealing, leeching, and reputation have been suggested as behaviours that could affect to what extent societies evolve moral behaviour (Spronck & Berendsen, 2009). Besides actions that affect an individual’s resources, ac-tions that affect the way a group deals with the available amount of food are worth studying. Hayden (1972) stated that hunter-gatherer societies regularly eliminated 5+% of a population through actions such as as killing, abortion, infanticide, senilicide and invalidicide. Motives for these actions could be effective management of sustenance; a society might not want to share their valuable resources with individuals that do not or can not con-tribute.

(22)

Furthermore, as the food constraint model showed, societies evolved to become part collaborative and part egoistic. Studying when and why agents perform an altruistic or co-operative action will contribute to understand-ing the effects of environmental drivers on moral behaviour. For example, it has been argued that altruistic punishment plays a fundamental role in stabilizing co-operation in hunter-gatherer societies (Sugden, 2012; Fehr & G¨achter, 2002). Therefore, individuals may choose to share when they are being watched by group members to avoid altruistic punishment. However, agents may decide to egoistically consume food when they believe there will be no repercussions. This will require the addition of a cognitive aspect, where agents are able to reason about the consequences of their actions on their individual well-being, as well as group well-being.

In regard to family relations, it would be more realistic if newborn chil-dren were cared for by parents or grandparents for the first dependent years of their life. In the current implementation, children immediately start inter-acting with their environment as if they were adults. However, in humans, children need their parents to provide for them and require substantial in-vestment during the first few years of their lives. Similarly, the elderly require their children to care for them in old age. Modelling these aspects of family bonds would result in a more accurate representation of nuclear families and family resource management. In addition, expanding family relations may provide valuable insights. As shown by the results of this research, family relations can drastically change the way a society behaves in certain environments. However, only nuclear family relations (father-mother-child) were implemented. Studying more complex family relations, such as child to parent and grandparent relations, sibling relations, mari-tal relations and complete family relations could be useful in understanding how humans evolved to be moral in their conduct towards others. Here, it would be interesting to study the distinction between the level of altru-ism towards family members and non-related members of society and what causes individuals to be more altruistic towards one or the other.

(23)

References

Bazzan, A. L., Bordini, R. H., & Campbell, J. A. (1999). Moral senti-ments in multi-agent systems. In Intelligent agents v: Agents theories, architectures, and languages (pp. 113–131). Springer.

Bazzan, A. L., Bordini, R. H., & Campbell, J. A. (2002). Evolution of agents with moral sentiments in an iterated prisoner’s dilemma exercise. In Game theory and decision theory in agent-based systems (pp. 43–64). Springer.

Choi, J.-K., & Bowles, S. (2007). The coevolution of parochial altruism and war. science, 318 (5850), 636–640.

Colson, E. (1979). The harvey lecture series. in good years and in bad: Food strategies of self-reliant societies. Journal of Anthropological Research, 18–29.

Cordain, L., Miller, J. B., Eaton, S. B., Mann, N., Holt, S. H., & Speth, J. D. (2000). Plant-animal subsistence ratios and macronutrient energy estimations in worldwide hunter-gatherer diets. The American journal of clinical nutrition, 71 (3), 682–692.

Darwin, C. (1871). The descent of man. Digireads. com Publishing. Falk, A., & Fischbacher, U. (2006). A theory of reciprocity. Games and

economic behavior , 54 (2), 293–315.

Fehr, E., & G¨achter, S. (2002). Altruistic punishment in humans. Nature, 415 (6868), 137–140.

Fenner, J. N. (2005). Cross-cultural estimation of the human generation interval for use in genetics-based population divergence studies. Amer-ican journal of physical anthropology, 128 (2), 415–423.

Gurven, M., & Kaplan, H. (2007). Longevity among hunter-gatherers: a cross-cultural examination. Population and Development Review , 33 (2), 321–365.

Hayden, B. (1972). Population control among hunter/gatherers. World Archaeology, 4 (2), 205–221.

Henrich, J. (2012). Social science: Hunter-gatherer cooperation. Nature, 481 (7382), 449–450.

Hill, K., Hurtado, A. M., & Walker, R. S. (2007). High adult mortal-ity among hiwi hunter-gatherers: implications for human evolution. Journal of Human Evolution, 52 (4), 443–454.

Jones, N. B. (1986). Bushman birth spacing: a test for optimal interbirth intervals. Ethology and Sociobiology, 7 (2), 91–105.

Kaplan, H., Hill, K., Lancaster, J., & Hurtado, A. M. (2000). A theory of human life history evolution: diet, intelligence, and longevity. Evolu-tionary Anthropology Issues News and Reviews, 9 (4), 156–185. Keeley, L. H. (1988). Hunter-gatherer economic complexity and

“popula-tion pressure”: A cross-cultural analysis. Journal of Anthropological Archaeology, 7 (4), 373–411.

(24)

Ruse, M. (1986). Evolutionary ethics: A phoenix arisen. Zygon, 21 (1), 95–112.

Schroeder, D. (2015, June 1). Evolutionary ethics. The Internet Encyclopedia of Philosophy. Retrieved from http://www.iep.utm.edu/evol-eth/

Smith, C. S. (2003). Hunter–gatherer mobility, storage, and houses in a marginal environment: an example from the mid-holocene of wyoming. Journal of Anthropological Archaeology, 22 (2), 162–189.

Spronck, P., & Berendsen, B. (2009). Evolutionary ethics in agent societies. International Journal of Social Robotics, 1 (3), 223–232.

Sugden, R. (2012). Altruistic punishment as an explanation of hunter-gatherer cooperation: How much has experimental economics achieved? Behavioral and Brain Sciences, 35 (01), 40–40.

Trivers, R. L. (1971). The evolution of reciprocal altruism. Quarterly review of biology, 35–57.

Whitley, D. (1994). A genetic algorithm tutorial. Statistics and computing, 4 (2), 65–85.

Wilensky, U. (1999). Netlogo. Center for Connected Learning and Computer-Based Modeling, Northwestern University. Evanston, IL. Retrieved from http://ccl.northwestern.edu/netlogo/

Drivers for the Evolution of Moral. Behaviour in Multi-Agent Societies