Rationality Within the Computer. The Myth of the Stupid Prehistoric Savages

Socio-ecological models make emphasis on physiological motivation, such as hunger, thirst, fatigue and comfort. In this case agents generate their goals around some physiological trigger, e.g., getting hungry.

In a sense, even computational agents implemented as biped stomachs can be considered “rational agents” because they make optimal decisions: they “want” to survive, and then they need to look for accessible resources. They have been programmed with the instinctive knowledge that they should hunt animals and gather for vegetables to acquire food, and therefore they hunt, gather and move looking for preys and resources. Janssen and Hill (Chap. 3, see also Janssen and Hill 2014) assume human hunting behavior is consistent with Optimal Foraging Theory, which is a model of animal behavior. In this way, hunter-gatherer foraging strategies—optimal group size, movement frequency and average daily return rate per hunter—are examined as the consequence of environmental factors—differ- ences in resource distributions—and not because of social or political dispositions.

Human (and even animal) rationality is much more complex than expected and therefore, it is easy to conclude that deterministic relationships between environ­mental stress and social change are inadequate (Mithen 1991; Costanza et al. 2007; Gardner 2012; Polechova and Barton 2015; Bryson 2015). The challenge of a computer simulation of human behavior is them to assess the impact of culture and knowledge on decision making behavior (An 2012).

We need to implement a form of intelligence beyond literal rationality if we want our historical models be credible. Socially intelligent agents (SIAs) should be defined as agents that do not only from an observer point of view behave socially but that are able to recognize and identify other agents and establish and maintain relationships to other agents (Dautenhahn 1998). The process of building SIAs will always been influenced by what the human as the designer considers “social,” and conversely, agent tools that are behaving socially can influence human conceptions of sociality. A cognitive technology (CT) approach toward designing SIAs would afford an opportunity to study the process of (1) how social agents can constrain their cognitive and social potential, and (2) how social agent technology and human (social) cognition can co-evolve and co-adapt and result in new forms of sociality. Aspects of human social psychology, e.g., storytelling, empathy, embodiment, and historical and ecological grounding, can contribute to a believable and cognitively well-balanced design of SIA technology in order to further the relationship between humans and agent tools.

One of the very first computer simulations of prehistoric hunter gatherers was that of Robert Reynolds (1986). He explicitly approached the problem of rationality in hunter-gatherer decision-making in terms of:

• the ability of each member to collect and process information about the resource distribution,

• the extent to which information is shared among members,

• the specific sets of decision available to each member, and

• the way in which the individual decisions are integrated to produce a group decision.

On that basis, Reynolds defined a general approach to programing that can also be considered as a general program for rationality in social evolution studies. He calls Cultural algorithm (CA) a specific kind of evolutionary computation frame­work where there is a knowledge component that is called the belief space in addition to the population component. The belief space of a cultural algorithm is divided into distinct categories representing different domains of knowledge that the population has of the search space. The belief space is updated after each iteration by the best individuals of the population. The best individuals can be selected using a fitness function that assesses the performance of each individual in population much like in genetic algorithms.

Reynolds lists different belief space categories:

• Normative knowledge: A collection of desirable value ranges for the individuals in the population component—e.g., acceptable behavior for the agents in population.

• Situational knowledge: Specific examples of important events—e.g., successful/unsuccessful solutions

• Temporal knowledge History of the search space—e.g., the temporal patterns of the search process

• Spatial knowledge Information about the topography of the search space

The “best-fitted” individuals of the population can update the belief space via an update function. Also, the knowledge categories of the belief space can affect the population component via an influence function. The influence function can affect population by altering the genome or the actions of the individuals.

The algorithm has been applied to find the optimum in a dynamic environment composed of mobile resources. The aim of this approach is to combine different knowledge sources to direct the decisions of the individual agents in solving optimization problems. Reynolds and collaborators developed an approach based on an analogy to the marginal value theorem in foraging theory to guide the integration of these different knowledge sources to direct the agent population (Reynolds et al.

Cultural Algorithms were developed by Reynolds as a computational framework in which to embed social learning in an evolutionary context. Unlike traditional learning approaches, Cultural Algorithms derive their power from large collections of interacting agents. Within virtual worlds it is often the case that we wish to coordinate the behavior of large groups of intelligent agents in an efficient fashion. Cultural Algorithms are able to perform large-scale group learning within these virtual worlds. They have been used to generate socially intelligent controllers and group social behavior in various simulated environments, both serious and fun.

Given that the study of differences between animal and human behavior emphasizes human motivation and purposefulness and it affirms that human behavior is shaped first and foremost by an intention held by the subject, any historical explanation based only on the idea of “adaptation” seems to be limited (Stutz 2012). The same criticism is applicable to traditional “rational-choice” explanation where each agent individually assesses its situation and makes deci­sions based on a fixed set of condition-action rules (Gulyas 2002). That makes many agent-based models nothing more than a discrete planning for expressing descriptions of intended courses of action. It seems as if some designer (be a computer scientist or a god) needs to know the society before modeling it (Grand 2012).

Humans act supposedly on the grounds of beliefs about world-states that they contribute to modify, and which will be modified by their actions. Consequently, the “cause” of any social action that may have occurred in the past lies in the agent motivations for performing it. Social actions have been defined in terms of pur­poseful changing of natural and social reality (Leont'ev 1974; Engestrom 1987; Wobcke 1998; Davydov 1999; Edwards 2000; Bedny and Karwowski 2004; Feldman and Orlikowski 2011; Thornton et al.

Goals, beliefs and intentions are in fact arbitrary interpretations of particular events (Bratman 1987).

Nevertheless, the frontier between intentional activity and operational behavior is blurred, and movements are possible in all directions. Intentions can be trans­formed in the course of an activity; they are not immutable structures. An activity can lose its motivation and become an action, and an action can become an operation when the goal changes. The motivation of some activity may become the goal of an activity, as a result of which the latter is transformed into some integral activity. Therefore, it is impossible to make a general classification of what an activity is, what an action is and so forth, because the definition depends on what the subject or object in a particular real situation is. The constitutive elements of a belief cannot be precisely separated in the same way that two actors can be isolated from one another. Even when we separate one actor from another, the fact that his or her beliefs depend to a great extent on previously acquired knowledge means that he/she cannot be completely separated from the environment in which such knowledge has been acquired.

An additional trouble is that social motivations have their own dynamics, often contradictory. In other words, social activities are not isolated entities; they are influenced by other activities and other changes in the environment. People interact, influence others, reinforce some actions, interfere with others, and even sometimes prevent the action of other people (Creary 1981). The term contradiction is used to indicate a misfit within the components of social action, that is, among subjects, needs, motivations, goals, actions and operations, and even mediating artifacts (division of labor, rules, institutions, etc.), and produces internal tensions in apparently irregular qualitative changes, due to the changing predominance of ones over others. Activities are virtually always in the process of working through contradictions, which manifest themselves as problems, ruptures, breakdowns, clashes, etc. They are accentuated by continuous transitions and transformations between subjects, needs, motivations, goals, behavior, signs, tools, rules, community, division of labor, and between the embedded hierarchical levels of collective motivation-driven activity, individual goal-driven action, and mechanical behavior driven by the tools and conditions of action. Here lies the true nature of social causality and the motivation force of change and development: there is a global tendency to resolve underlying tension and contradictions by means of change and transformation. Since social activity is not relative to one individual but to a distributed collection of interacting people and the consequences of their actions, we cannot study how social activities took place by understanding the intentions or motivations of individual agents alone, no matter how detailed the knowledge of those individuals might be. To capture the teleological or purposive aspect of behavior, we should investigate collective action, that is, why different people made the same action, or different actions at the same place and at the same time. Its research goal should be to explain the sources or causes of that variability, and not exactly the inner intentions of individual action.

What we need to study is the constant interaction between agent and context. Consequently, the basic unit of motivation is not the discovery of some verbal proposition such as “x believes that P”, “x desires that P”, “x knows that P”, and so forth. Rather, we are aware of those things that are playing a prominent role in constraining the global constraint satisfaction settling process in our minds (O’Reilly and Munakata 2002, p. 218). What constitutes “causality” is not just the “consciousness” of the reasoning system itself, but also the rich matrix of relations it bears to other agents, practices and institutions.

Beliefs, desires and intentions may change according to the variation in the local conditions, and according to what the agent may “learn” from its environment and from other agents interacting with it. To a certain extent, this is just giving the agents another level of rules. However, the nature of these rules is different, in that they are meta-rules about how to form rules (Lee and Lacey 2003). Such meta-rules allow for the type of self-reference that is key to the historical explanation. Individual social actors, in going about their various interactions, form representations of those interactions. Moreover, they abstract away from the details of individual interactions to formulate underlying rules that describe these interactions. In the case of societies, these rules are usually called norms or institutions. When individuals form abstractions about what the normative behaviors are in their society, they begin to act on them, either by behaving differently themselves or by reacting differently to the behavior of others. Thus, the abstractions the individuals make drive the behavior that emerges from their interactions. But the loop does not end here. As the indi­viduals continue to interact based on the abstractions they have made and new societal patterns emerge, the individuals will make abstractions about those new patterns, which in turn will give rise to a new set of emergent behavior, ad infinitum (Baumer and Tomlinson 2006). Computational agents should be designed as learning entities, gaining ever more accurate information from the effects of their actions or more successful strategies from observing others’ behaviors. In substance, reinforcement learning is shaped on the model of evolution, a fitness formula being always implied. This essentially has led to implement agents’ capacity to gain more accurate information. But this is only part of the job required by a dynamic model of agent-hood. Agents undergo social influence, come to share the same beliefs and expectations, squeeze into the same practices, and this type of social influence sometimes leads them to form inaccurate, even wrong beliefs.

As a result of this focus on social actions as practiced by human actors in reference to other human actors, the idea of “agency” appears to be synonymous with an agent's way of being, seeing and responding in the world. It is an embedded and interpreting agency that draws on its funds of knowledge to both interpret and respond to the environment (Edwards 2000). The agent interprets and responds to the contexts of action and exploits the opportunities for effective action within them. It is an outward-looking mind which seeks local scaffolding to enhance its purposive action. However, it is not possible to fully understand how people act and work if the unit of study is the unaided individual with no access to other people. The unit of analysis is object-oriented action mediated by human produced tools and signs. Thus, we are constrained to study context to understand relations among agents, actions and goals. This is the reason of emphasizing the use of invariant change-relating capabilities to characterize historical events. What humans did and the way they did it is firmly and inextricably embedded in the social matrix of which every person is a member.

The obvious conclusion is that we are far from being perfect rational agents. And our ancestors even less, given the poor access to information at real time to take decisions (Leaf 2008). Nevertheless, we are very far from the current “stupidity” of usual computational agents in their eternal search for optimal but simple solutions.

Oestmo et al. agents' (Chap. 4) appear to be a bit more rational and less “adaptive” than traditional “optimal foragers”. They know that they should produce tools to increase the possibilities of having success in hunting and gathering. This awareness on the necessity of technology is what makes them more “human” than strictly animal. They look for explanations of the changing raw material preferences when deciding to make tools. From a naturalistic point of view explanations for change in human raw material usage frequency would only include climate/environmental change and its co-variability with mobility and procurement strategies, the selection of certain raw materials for their physical properties, changes in demography, etc. All these parameters are well within a strong and limited rationality hypothesis, where prehistoric people, made optimal choices. But Oestmo et al. also consider the preference for appearance or color, symbolic value, and style. The authors have created a simple model of one forager with a mobile toolkit of fixed capacity that is randomly placed on the environment. The simulation is based on a previous model by Brantingham (2003). This “stupid but rational” behavior is compared to archaeological data from prehistoric settlements around the town of Mossel Bay, Western Cape, South Africa, offering a long sequence of change in raw material selection. Here, optimal decisions are not assumed, but tested against rel­evant prehistoric data. They think this is the right approach, instead of assuming the truth likeness of a particular theory of the way people made choices in a world with low developed means of production; we should look for particular tests of the explanatory power of the hypothesis. Oestmo et al. results should be compared with similar work by Davies et al. (2015), Clarkson et al. (2015), Pop (2015).

O’Brien and Bergh's agents (Chap. 6) are apparently even more “human”, in the sense that the agents in their virtual world can build their own complex cognitive map of the environment around them, and they can make reference to ideological, social and political factor to motivate their decisions.

There is a growing interest in the computer science and artificial intelligence community to build more credible virtual agents that may act in a simulated world in the same way we believe humans would have acted. The belief-desire-intention software model (usually referred to simply, but ambiguously, as BDI, see Rao and Georgeff 1998; Wooldridge 2000; Luck et al. 2004; Bosse et al. 2011; Caballero et al. 2011; Taillandier et al. 2012; Kennedy 2012; Kim et al. 2013; Gelfond and Kahl 2014; Blount et al. 2014; Pantelis et al. 2016) is a software model developed for programming intelligent agents. In essence, it provides a mechanism for sepa­rating the activity of selecting a plan (from a plan library or an external planner application) from the execution of currently active plans. Consequently, BDI agents are able to balance the time spent on deliberating about plans (choosing what to do) and executing those plans (doing it). A third activity, creating the plans in the first place (planning), is not within the scope of the model, and is left to the system designer and programmer.

Alternative virtual agent architecture is “PECS” (Urban and Schmidt 2001; Malleson 2012) which stands for “Physical conditions, Emotional states, Cognitive capabilities and Social status”. The authors of the architecture propose that it is possible to model the entire range of human behavior by modelling those four factors. PECS is seen as an improvement over BDI because it does not assume rational decision making and is not restricted to the factors of beliefs, desires and intentions (Schmidt 2000). Instead, an agent has a number of competing motives (such as “clean the house”, “eat food”, “raise children”, “sleep” etc.) of which the strongest ultimately drives the agent’s current behavior. Motives depend on the agent’s internal state (an agent with a low energy level might feel hungry) as well as other external factors (an agent who smells cooking food might become hungry even if they do not have low energy levels). Personal preferences can also come into play, where some people feel a need more strongly than others even though their internal state variable levels are the same (Balke and Gilbert 2014). In this sense, Ho et al. (2006) have proposed a Categorized Long-term Autobiographic Memory (CLTM) architecture, utilizing abstracted notions of human autobio­graphic memory and narrative structure humans apply to their life stories (see also Pointeau et al. 2013; Lei et al. 2013; Boloni 2014). El-Nasr et al. (2000) and Resisenzein et al. (2013) have explored the simulation of human emotions.

H-CogAFF cognitive architecture gives place to emotions and other high level cognitive layers integrating “fuzzy” boundaries between different levels of func­tionality, and allowing for some of the information-processing mechanisms to straddle two or more layers (Sloman and Christel 2003; Sloman 2011; Petters 2014; Goertzel et al. 2014).

Other recent essays in this direction are Goertzel et al. (2013), McRorie et al. (2012), Gratch et al. (2013), Faur et al. (2013), Kang and Tan (2013), Anastassakis and Panayiotopoulos (2014), Haubrich et al. (2015). For more advanced issues, see

Bostrom (2012), Hughes et al. (2012), Schonbrodt and Asendorpf (2011). Current work on cognition and artificial intelligence will allow the proper understanding of motivation in social action (Pollock 1995; Friedenberg 2011; Tenenbaum et al. 2011; Chella and Manzotti 2013; Diettrich 2014; Bryson 2015; Campenni 2016).

1.2.4