The function perspectives of theSurvivability Frameworkcan be used to understand and compare existing system architectural approaches, in the same manner that theCogAff schematic framework (Sloman, 2002) is used to understand the structures of existing architectures. Such comparisons are made possible by distinguishing the components in each architecture that falls into each functional perspective. These comparisons are meant to highlight the structural differences between different architectural ap- proaches, especially their advantages and weaknesses. The functional perspectives of the framework form a classification scheme for existing architectures.
Table 4.5: Examples of arbitration schemes for selection of actions.
Mechanism Description
(a) Generic arbitrator
Selection desired action(s)
a1 a2 a3 : aN
Selection desired action(s)
a1 a2 a3 : aN
An arbitration or selection process in which an action is selected from among several actions, based on their relative activation levels.
(b) Weighted vector summation
Σ desired action(s)
a1 a2 a3 : aN
Σ desired action(s)
a1 a2 a3 : aN
The summation of vector outputs of each action, weighted by their activation values. Also known as command fusion, this is a cooperative scheme, as no actions are being suppressed in favour of others, and all actions contribute cooperatively to the final output vector.
(c) Competitive (Greedy)
desired action(s)
a1 a2 a3 : aN
{..}
max argα∈Α
desired action(s)
a1 a2 a3 : aN
{..}
max argα∈Α
The greedy selection of the action for which the activation level is the highest. This simple strategy results in a winner-takes-all situation, as other actions whose activation levels are below the maximum, would have no contribution to the final outcome. Consequently, this scheme is useful for competing actions which must not be executed at the same time.
(d) Finite State Machine
FSM desired
action(s)
a1 a2 a3 : aN
FSM desired
action(s)
a1 a2 a3 : aN
The use of a finite state machine (FSM) is to sequence actions which have temporal or ordering dependencies, i.e. if certain actions need to precede others, and where the transitions between actions are dependent on the occurrence of events. Implemented as a state machine, e.g. petri-nets, or with scripts.
To demonstrate its viability as a classification scheme, a comparison of architectural approaches using theSurvivability Frameworkis shown in Table 4.6. Each architecture is a known approach for the solution of the autonomous agency problem (section 2.4.1).
Approaches (a)–(f) are architectural approaches for the control of mobile robots, and autonomous systems, while (g)–(i) are cognitive approaches for understanding human information processing, as discussed in sections 3.5 and 3.7.2.
The framework differs from behaviour-based (Arkin, 1998) and task-oriented ap- proaches in several ways, as shown in Table 4.7. In task-oriented, behaviour-based approaches, decisions are made based on the tasks of a system, where behaviours are
Table 4.6: Comparing architectural approaches using the Survivability Framework.
Approach ENVIRONMENT PREFERENCES MECHANISMS CAPABILITIES
(a) Subsumption (Brooks, 1991)
No world model;
Direct sensing.
Activations;
Inhibitions.
Subsumption. Reflexive actions;
Reactive.
(b) Behaviour Based (Arkin, 1998)
No world model;
Direct sensing.
Activations;
Inhibitions.
e.g. Finite State Machine (FSM), Motor Schema, Potential Fields.
Reactive behaviours.
(c) Top-Down;
Hierarchical; e.g.
(Alami et al., 1998)
Model-based. Goals and Objectives.
e.g. FSM, Motion Planning.
Planned actions.
(d) Probabilistic;
e.g.
(Sebastian Thrun, 2005)
Model-based, Probabilistic.
Probability Distribution Functions.
e.g. Bayesian Inference;
Dempster-Shafer Theory.
Deliberative actions.
(e) 4D-RCS Reference Architecture (Albus et al., 2002)
States and Observations;
World Model.
Value Judgements.
Behaviour Generation.
Planned behaviours.
(f) DAMN Architecture (Rosenblatt, 1997)
Model-based;
Sensor Fusion.
Votes issued by each behaviour module.
Arbitration (voting) of behaviours.
Distributed behaviours.
(g) EMIB Architecture (Michaud, 1995)
States and observations.
Motivational drives and fuzzy emotion variables.
Intentional behaviour generation and selection
Motivated behaviours.
(h) SRK Model (Rasmussen, 1983)
Symbols, Signs, Signals.
Activations;
Inhibitions.
Planning, Rule-matching, Sensory-motor couplings.
Knowledge-, Rule-, Skill-based behaviours.
(i) Affective (Norman, Ortony, and Russell, 2003)
Direct sensing to derive affective states.
Affective states i.e. Emotions.
3-layer processing at thereflection, routineand reactionlevels.
Reflective, Routine, Reactive behaviours.
(j) Survivability Framework i.e.
(Quek et al., 2006a)
Model-based, Needs-model.
Needs, Emotions.
Maximization of needs-fulfilment and positive emotions.
Actions that satisfy needs.
Table 4.7: Comparison of Survivability Framework and task-oriented approaches.
Task-oriented approaches Survivability Framework (a) Decisions made based on which tasks
need to be performed.
(a) Decisions made based on needs, which incorporates tasks to be performed, not just on tasks alone.
(b) Behaviours are designed with the aim of decomposing complex tasks into smaller portions.
(b) Behaviours are designed to specifically fulfillneeds, not only to accomplish tasks.
(c) Do not directly address necessary conditions for survival.
(c) Directly addresses necessary conditions for survival.
(d) Behaviours are selected depending on heuristics /a prioripolicy / cost-function.
(d) Behaviours are selected in order to maximizeneedsfulfilment.
designed with the aim of decomposing complex tasks into smaller sub-tasks which can be accomplished by a system’s collection of actions. The framework accomplishes this and more, by basing its decisions on theneedsof a system, in addition to the tasks. In this framework, behaviours are designed to specifically address the fulfilment of theneeds of a system. In terms of behaviour selection, conventional methods rely on heuristics or somea prioripolicy or cost functions; this is not the case in the framework, in which a system’sneedsandemotionsare the drivers for the generation and selection of behaviour that directly maintains theirwell-being.