• No results found

Convention research in other fields

4.3 Perspectives on conventions

4.3.3 Convention research in other fields

There is a variety of research in other fields relating to convention emergence, particularly in the psychological and economic domains. Human populations are particularly adept at efficiently emerging conventions without explicit agree- ment, and incorporating insights from these fields into the theory of conventions typically used in agents research may yield significant benefits. In this section, we focus on the work of Garrod (1994), whose experiments with humans pro- vides significant insight into how convention emergence occurs in our society, and Boyer and Orlean (1992), who discuss convention behaviour and manipula- tion in the latter stages of the convention lifecycle. The latter is a particularly under-investigated area in open MAS. We use the results from these works in the development of our definition and metrics of convention, and in investigating how to manipulate convention emergence in the latter stages of the convention lifecycle (Chapter 7) respectively.

Garrod (1994) investigated human convention emergence, using an exper- iment in which people evolve coordinated description languages when solving problems. Volunteers are paired and play a maze game in which they must describe maze positions to the other player in order to prevail. The volunteers were divided into two groups: one in which individuals remained in the same pair for the duration of the experiment (9 iterations of the maze game), while the other group changed pairing each game (within the community represented by the group). Analysis of the language used to describe the maze shows three

categories of representation used by pairs: (i)line, such as “first row and third

column”, (ii)path, such as “two along from you”, and (iii)matrix, in which verti-

cal and horizontal lines are named according to a scheme such as that from chess (e.g. “C4”). Within each category, there are a number of different instantiations

of representation.

Pairs in the community group took longer to agree than the isolated pairs. However, their representation (a form of matrix) was agreed across the entire community, despite communication only being local between pairs. Conversely, isolated pairs quickly agreed on a representation, but the representations were wildly divergent between pairs. From the experiment results, Garrod (1994) concluded the following.

1. There is a local coordination process between individuals in a pair, by which local precedence and salience have a strong influence on the repre- sentation that is chosen.

2. When the community has agreed on a representation, players have a stronger preference to act according to convention than to the constraints of local precedence and salience.

3. When two players meet with conflicting chosen representations, the rep- resentation that tends to get chosen (mostly implicitly) is that which is

represented most frequently in thejoint history of the players.

4. The community group explores potential representations more widely at the beginning than the isolated pairs group, before settling on a single representation across the entire community.

5. The local coordination process between individuals, which incorporates precedence and salience, leads to a global coordination process as those individuals interact with others in the population and “infect” them with their preferences.

6. An important component of the community convention emergence is the ability of individuals to monitor their communicative success — typically modelled in the agents community through reinforcement learning and/or a personal interaction history.

7. High simulated population churn, achieved by pairing individuals with others in such a way that they do not form a community with repeated

interactions, decreased the levels of coordination drastically. This has

significant implications for open MAS with high levels of population churn. Garrod (1994) argued that the results are a consequence of community effects rather than the influence of a given individual, since there is not sufficient time in the experiments for an individual to gain influence or standing within the group. However, the influence of an individual over a population is complex with respect to conventions, since precedence creates feedback effects, and any one action might propagate significant changes through a population. As such, it is not appropriate to entirely disregard the potential influence of individuals in convention emergence, although the results suggest that community effects can transcend individual actions. We use the influence of individuals as the basis

of our Influencer Agent mechanism, which is explored in subsequent chapters.

The setup used by Garrod is equivalent to a population connected by a fully- connected network, which also would not imbue any individual with special

standing from network effects2.

There are relatively few investigations into the latter stages of convention

establishment. One of the most detailed was presented by Boyer and Or-

lean (1992), who attempt to account for how an established convention can be superseded by a preferable one. They identify four situations that can overcome the reinforcement feedback effects of an established convention.

1. Convention collapse. If the environment (including the population of

agents) is drastically and suddenly altered, the previously established con- vention may lose its force of precedence and provide room for new con- ventions to become established.

2. External invasion. When a new group with an alternate convention joins a

population, the force of precedence already present in the new group allows 2We investigate this type of special standing in Chapter 6.

the alternate convention to gradually undermine the currently established one.

3. Translation. If there is a possible new convention that is compatible with

the previous convention, then the costs associated with adopting the new convention are removed and the new convention can become established more easily.

4. Collective agreement. If a sufficient proportion of the population explicitly

agrees to collectively adopt a new convention it can easily become estab- lished. Boyer and Orlean (1992) note that this requires the existence of a central authority, and since this is impractical for open MAS we do not consider this route to convention change.

Boyer and Orlean (1992) also conclude that the best way for a new con- vention to become established is to begin in a localised space and progressively invade the rest of the population. This is a form of “gradual accretion of prece- dence” (Young, 1993).

There has been little research into co-existing sub-conventions, with the

notable exception of Villatoro and Sabater-Mir (2011) who propose social in-

struments as a way of identifying and mitigating less desirable sub-conventions,

that might be supported by self-reinforcing underlying network structures. How- ever, they assume that sub-conventions are necessarily those that we want to destabilise, and do not consider the potential desire to support co-existing sub- conventions.

The contributions of Boyer and Orlean (1992) and Villatoro and Sabater- Mir (2011) are concerned with how conventions might be manipulated by in- terested parties in order to support the adoption of desirable behaviour across

the population. However, research in this area is limited and there are, to

our knowledge, no proposed mechanisms for manipulating which convention a population adopts in open MAS. Such a mechanism is the focus of Chapter 5.