The role of Attribures in Mental State Multi-Agent Architectures

The mental state framework (MSF) is suited to build agents architectures following the mental state paradigm. It encloses three different main components that provide a set of core attributes, a set of laws and a set of controls. Those components allow the definition of rules that guide the agents' behaviour.In this dissertation I discuss this paradigm in depth, by introducing two different implementations under two domains: Robosoccer and Pursuit, in which I study the roles of attributes.In MSF, attributes are essential elements, being present ever through the agents' reasoning process, from perception to action execution.

MSF

In MSF, each mental state keeps information about external world and also information about their internal state variables (for example, variables which deal with energy and memory management). That information is embedded into agents' mind, and is passed to attributes through an evaluation process. Attributes such as insistence, uncertainty, importance, intensity, dissatisfaction and urgency have specific roles in the architecture. They decide about the importance, the uncertainty, etc., of a certain mental state, i.e., they evaluate each mental state in different dimensions.

Role of attributes

For example, dissatisfaction is related to mental state desire and determines the degree of willingness to change a state of the world in order to fulfil that desire, whilst importance, under the same mental state, measures the relative priority an agent puts in pursuing that desire.Moreover, mental states have a state that can be active or inactive. For example, usually all beliefs are active whereas intentions when active will be responsible for the agents' actions. Changing the state of all mental states in an agents' mind provides the system with a dynamic transformation that will ultimately characterise agents' behaviours.Rules are created based on laws and controls. Those rules are responsible for mental states activation.

BDI rule example

For example, in a BDI-like architecture, a rule of the type B + D -->I, will activate I while B and D conditions trigger the rule.In the Robosoccer environment, I studied how attributes could be used to control agents' behaviours. The issues discussed under this initial experimental scenario included the evaluation of the role of each attribute for triggering rules, the inquisition about the way intentions were selected while in presence of contradictions, and finally, the test about how attributes could be selected in order to control the duration of the activity of mental states, specifically intentions. This was an initial step that drove me to wonder about how attributes as a whole could be responsible for planning execution activity.

As a matter of fact, attributes within each mental state carry to mental states a multi-dimensional perspective, which ultimately converges to intentions that automatically acquire a strength derived from a desire or a value of importance supported by a certain belief. In my point of view, this multi-dimensional perspective should be carried into planning execution, because the richness of information inside each one of the attributes should be visible in terms of action.So the question was to evaluate how could attributes change behaviours inside a plan, executed with respect to an intention?This question was the primary concern throughout all the second part of this research.

PLANNING USING AND-OR TREE STRUCTURE

I decided to implement an AND-OR tree to support agent planning. Agents have now to choose a (sub-)goal from a partial plan, supported by a set of beliefs that constraint agents choices. In fact, I had the opportunity to present, step by step, a set of experiments in the Pursuit environment (a simpler scenario than Robosoccer) where behaviours were controlled by attributes.In the end it became clear that attributes can help agents to control available resources, to manage deadlines, to tune behaviour based on goals' persistence, and also, to manage priorities. I claim that in an unpredictable and dynamic environment this can be an advantage.In the end I grasp a set of guidelines that could be helpful to endorse a (proto-)theory about how, in a general sense, attributes could be used to endow personality characteristics to agents. I explore this idea and make a small experiment to support a mainly theoretical definition of what I call the agent's personality.

QUESTIONS

Nevertheless some questions deserve a deeper investigation. Some of them are the following:

- How could one measure the `usefulness', in terms of multiagent systems, of ascribing attributes to agents?

- What could be the impact in a multiagent system of agents with personality as they are described in this work?

- How can agents behaviour be tuned by changing values of their attributes ?

CONTRIBUTIONS

I summarise the original contributions of this work as follows:

1- Discuss the role of attributes in MSF framework architectures, as fundamental elements:

a- To explain how they enrich the way information is perceived in the environment by evaluating it through different dimensions based on attributes evaluation (each one attached to different mental states) .

b- To endow agents with tools to select goals based on different attributes dimensions.

2- Show how attributes are present through all MSF architecture and, in particular, to explain how they participate in path options and co nstraint selection within a (partial) planning tree.

3- Discuss agents' personality based on attributes, related to how attributes link to each other and also how they ascribe behaviours.

2008-05-19 - José Cascalho