A Study of Agent Influence in Nested Agent Interactions

Logie, Robert Davin (2010). A Study of Agent Influence in Nested Agent Interactions. PhD thesis The Open University.

DOI: https://doi.org/10.21954/ou.ro.0000f247


This work develops a theory of agent influence and applies it to a coached system of simple reactive agents. Our notion of influence is intended to describe agent ability which is contingent on the actions of other agents and we view such behaviours as being “nested”. An agent may have the ability to make A hold only if another agent has carried out a particular action. Our analysis of this is based on a combination of the observation of the effects of an agent’s actions in a bounded environment and observations on what may be changed in that environment and is intended to allow for a logical representation of nested behaviours. We build on this notion to develop a theory of influence which we offer as an extension of existing systems for representing agency and its effects.

The notion of an agent being able to “see to it” that something is brought about has been a useful device for reasoning about agent ability. These so-called STIT semantics have been developed by a number of researchers. Standard STIT semantics allow statements of the form [α stit: A] which says that agent a has the ability to see to it that A holds. Although based on the concept of agent action STIT semantics also allow for the representation of concepts involving what may be thought of as inaction. An agent deciding, for example, not to execute a particular action may be characterised as seeing to it that it does not see to it that A, [α stit: [α stit: -A]]. STIT encourages nesting and although this nesting extends across actions within an agent it does not extend easily across agents. So called other agent statements of the form [β stit: [α stit: A]] do not make sense in standard stit semantics because β seeing to it that α sees to it that A holds implies that β has some dominion over a which, in turn, compromises α’s agency. Although the statement makes no sense under standard STIT it does make sense in an intuitive way and Brian Chellas [31] notes that it would be:

“...bizarre to deny that an agent should be able to see to it that another agent sees to something”

This is also mentioned in Belnap et al. [8, page 275]. Chellas is correct and there are numerous settings in which other agent STIT does make sense. These settings, which are captured in various readings of STIT, may bring a great deal of system level overhead. In a normative system, for example, β may have the option of imposing a sanction on α if α fails to bring about A and in this sense may be thought of as seeing to it that α sees to it that A holds. Similarly a deontic reading may place β in a position where it is able to place an obligation on α to bring about A. These readings allow for sensible interpretation of other agent STIT but the examples above require that agents have sufficient awareness of personal utility be able to manage sanctions or that they are able to reason about obligations. These readings offer nothing for simple agents with limited resources and abilities.

We offer another reading for the STIT element, one based on the concept of agent influence and one which carries minimal system level overhead. Because influence may be contingent on simultaneous or sequential behaviour by a number of agents it is extendible across agents and offers a means of addressing other agent statements. We extend the standard STIT semantics of Horty, Belnap and others with the introduction of “leads to” and “may lead to” operators which allow us to move our analysis into a setting where observation provides evidence of influence. We then explore the manifestation of influence in a number of scenarios. After exploring how influence manifests itself we then offer a partial logical characterisation of the influence operators and discuss its relationship with standard STIT.

Building on these semantics and the partial logical characterisation we then explore the practical use of our theory of influence in an agent learning system. We describe experiments with a system specified by safety and liveness properties and having two broad classes of agents, actors and coaches. Actor agents will manipulate their environment and coaching agents will observe the actor’s behaviour and its effects using aggregated observations to generate new behaviours which are then seeded in the environment to modify actor behaviour.

We then offer a discussion and evaluation of our theory and its applications indicating where it may be further developed and applied.

Viewing alternatives

Download history


Public Attention

Altmetrics from Altmetric

Number of Citations

Citations from Dimensions

Item Actions