Chapter 3 Foundations of Survivability in Autonomous Systems 55
3.7 Cognitive basis for survivability
3.7.2 Human decision making and information processing
The human decision making process can be studied by examining the manner in- formation is being processed. Many models of human behaviour and information processing adopt a three-stage, three-layer configuration. The first partitions these into input-processing-output (IPO) stages typical of artificial systems; the second classifies them into layers representing processing with different levels of sophistication and granularity of information. Both configurations are alternative ways of partitioning the architecture of an organism (Sloman, 2002).
SENSORY REGISTER
WORKING MEMORY
LONG-TERM MEMORY Thought / Decision Making
Perception Response
Selection
Response Execution ATTENTION
RESOURCES
PROPRIOCEPTIVE FEEDBACK
EXTEROCEPTIVE FEEDBACK
PERCEPTUAL ENCODING CENTRAL PROCESSING RESPONDING
Figure 3.4: A model of human information processing. This is adapted from (Wickens et al., 2004), with the inclusion of two feedback paths,proprioceptiveandexteroceptive.
Stages in human information processing
There are three stages in the human information processing system, namely the perceptionorperceptual encodingstage, thecentral processingor central executive stage, and theactionor respondingstage (Sloman et al., 2005; Wickens et al., 2004). Figure 3.4 depicts this three step process. As shown in this figure, the sensory register receives input from the different senses of the body, and these are perceived. Perception recalls information from long-term memory for recognition, correlation, and pattern matching.
The end result is a set of percepts and internal states that are used in decision making, i.e. at the central processing stage. Here, the information generated by perception is evaluated for decision making on appropriate actions to be taken. These partial decisions and percepts are temporarily stored in global working memory. Once a response is selected, it is executed.
The operation of all these processes in parallel is influenced by the amount of attentional resources allocated to each. Attention is selective, in order to focus on the sensory information, decisions, and responses that are more relevant at any particular time. Finally, feedback on the outcome of the response is obtained both internally (proprioceptive) and from the environment (exteroceptive). Together, these processes
Motor Sensory
Reflective Level
Routine Level
Reactive Level
Figure 3.5: A three-level model of human behaviour (Norman et al., 2003). Processing at each level serves two functions: (i) evaluation of world, i.e. affect; and (ii) interpretation of what is happening in the world, i.e.cognition.
form a state feedback control system, and hence, it may be possible to realize this model with the use of control theory, subject to the completeness in replicating these processes in an artificial manner. An attempt to do this is shown in section 4.8.3.
Models of human decision making and effective functioning
A three-level configuration is prevalent in many human information processing models, namely those of Ortony et al. (2005), Norman et al. (2003), Rasmussen (1983) and Wickens et al. (2004). The three-layer model by Ortony et al. (2005) (shown in Figure 3.5) comprises affective and cognitive processes at thereactive,routine, andreflectivelevels.
This is analogous to the automatic, intuitive and analytical levels in Wickens et al.
(2004)’s model of adaptive decision making (as illustrated in Figure 3.6), and Sloman (2002)’sreactive mechanisms, deliberative reasoning, and meta-management layers(as shown in Figure 3.7). An ontological comparison of these information processing models is depicted in Table 3.2, showing the consistency of the three levels across different models, and with the types of behaviours dominant in each level. These behaviours are termed skill-based, rule-based or knowledge-based, according to the Skill-Rule- Knowledge (SRK) model (Rasmussen, 1983). Reconciliation of these models with the three-stage IPO model of information processing is made possible by the realization that these processes in fact, refer to the different levels of processing on-going within the central executive that interact with sensory receptors and motor effectors of varying complexity. Such is indeed the case in theCogAff schematic framework (Sloman, 2001;
Sloman and Chrisley, 2005), shown in Figure 3.7. To avoid taxonomical confusion, the levels would be referred to as the reaction/reactive, routine, and reflection/reflective, without loss of generality in meanings associated with each of these levels in other models shown in Table 3.2. (For instance, thereflective level would be taken to mean
EVALUATE EXPLANATION Run mental simulations Search for data Check consistency
EVALUATE ACTION
Run mental simulations Consider outcome, costs, and benefits
EVALUATE PLAN
Run mental simulations Consider timeline, and effectiveness
(Diagnosis) and Identify System
State
Assess Goal and Define Overall Task
Develop Plan
(Cue Integration)
Execute Action(s) Perceive Cues
Retrieve Cue-Action
Rule(s) LEVEL 2
SITUATION AWARENESS
LEVEL 1
SITUATION AWARENESS
LEVEL 3 SITUATION AWARENESS (Anticipation)
Monitor Effects of Actions
Track step completion Track context, purpose Anticipate need for action Monitor effect on system Feedback
SELECTIVE ATTENTION
WORKING MEMORY
Diagnosis Choice Action ANALYTICAL (Knowledge-Based)INTUITIVE (Rule-based)AUTOMATIC (Skill-based)
SCHEMATA PLANNING NETS
Cue patterns
Causal relationships MENTAL MODELS Hypotheses
Goals and Actions CUE-ACTION RULES Risk and values
LONG-TERM MEMORY
Figure 3.6: An integrated model for adaptive decision making (Wickens et al., 2004), illustrating the operation of three levels of information processing, selective attention, and how situational awareness arises at different levels.
Perception Central Processing Action
Meta-management (reflective processes)
Deliberative Reasoning (“what if” mechanisms)
Reactive mechanisms
Figure 3.7: The CogAff schematic framework (Sloman and Chrisley, 2005), used for explaining the different types of architectures that emerge from nature, and those of virtual machines.
all of theanalytical,meta-management,knowledge-basedbehaviour processing content present at that level).
A comparison between the different levels is shown in Table 3.3, where each level is elaborated by its features. Thereactivelevel is evolutionarily the oldest; it is responsible for automatic, skill-based information processing that generates the reflexive motor
Table 3.2: An ontological comparison of information processing models.
Level Ortony et al. Wickens et al. Sloman Rasmussen
(2005) (2004) (2002) (1983)
upper reflective analytical meta-management knowledge-based middle routine intuitive deliberative reasoning rule-based lowest reactive automatic reactive mechanisms skill-based
Table 3.3: Comparison between different information processing levels.
Level of information Evolutionary Demand on Representation Nature of processing development attention of Information Processing reflective, analytical most recent high symbols anticipation, planning routine, intuitive recent moderate signs rule-matching reactive, automatic least recent minimal signals sensory-motor
reactions to direct stimuli which are genetically inherited orwiredinto the organism.
This level monitors the environment through rapid perception of sensory signals with minimal processing. Under the influence of crisis, processes at this level interrupts higher-level processes to regain attentional resources that heighten arousal, and hasten response (Norman et al., 2003). This distinguishes it from theroutinelevel where the behaviour is conditioned from past learning. Theroutinelevel is an evolutionarily more advanced level of information processing responsible forintuitive,rule-based processing that lead to patterns of behaviour that are learned, or conditioned from past learning.
This is in the form of "if-then" rules which are formed from associations between cues or signs obtained from the senses and appropriate actions taken.
Tasks which require more than the simple selection and execution of rules are processed at the reflective level. This is the level which distinguishes human beings from most other organisms, as it is where higher-order cognition is located, and which is responsible foranalytical, knowledge-based processing that heavily relies on mental simulation to generate plans for execution (Wickens et al., 2004). Reflection is a meta- process (Norman et al., 2003) where the mind deliberates about itself, performing operations on its own internal representations of the situation and with awareness of its embodiment (i.e. the notion of theself), and its collection of capabilities, behaviours and the environment. This is the level that is activated whenever a novel situation is being encountered for the first time. With subsequent occurrence of a previously-
encountered situation, the response to the situation may be delegated to a lower level of processing i.e. the routine level, since experts at performing a specific task require less mental resources than say, a complete novice at performing the same task (which may even require his full mental capabilities), until the point when proficiency is achieved (i.e. when a set ofruleshave been learned to help perform the same task at a lower level subsequently).
The above models of human information processing can be reconciled with neu- roanatomical accounts of processes occurring in the nervous system. Ortony et al.
(2005) observe that the three levels, namely the reactive, routine and reflective levels, corresponds approximately to the spine/midbrain basal ganglia, cortex, cerebellum and prefrontal cortex, which not only attests to their differences in evolutionary development, but also suggests that affect, emotion and motivation are an intricate part of information processing at various levels.