Affect and machine design: lessons for the development of autonomous machines.

Animals and humans have two distinct kinds of information processing mechanisms: affect and cognition. Cognitive mechanisms--mechanisms that interpret, understand, reflect upon, and remember things about the world--are reasonably well understood. But there is a second set of mechanisms, equally important and inseparable--the system of affect and emotion that rapidly evaluates events to provide an initial assessment of their valence or overall value with respect to the person: positive or negative, good or bad, safe or dangerous, hospitable or harmful, desirable or undesirable, and so on.

Although affect and cognition are conceptually and to some degree neuroanatomically distinct systems, from a functional perspective they are normally deeply intertwined. They are parallel processing systems that require one another for optimal functioning of the organism. There is some evidence (1) that people with neurological damage compromising their emotional (affective) systems become seriously limited in their ability to organize their day-to-day lives, even while appearing to perform normally on a battery of standardized cognitive tasks. They become ineffective actors in a complex world. Furthermore, psychologists and others interested in artificial intelligence have repeatedly urged that affect is essential for intelligent behavior (2) by altering goal priorities and generating interrupts (e.g., References 3-5).

This paper (6) is intended to start a discussion about how the study of affect in biological systems might contribute to the development of autonomous computer systems. We suspect that from a functional perspective, some of the evolutionary forces that presumably led to the emergence of affect in animals are likely to be relevant to the design of artificial systems. However, we view this paper as only setting the stage for further research, realizing full well that it raises many more questions than it answers.

A model of affect and cognition: Three levels of behavior

In this section we outline the essence of our three-level theory of human behavior, a work that is still in progress, (7) after which we discuss how these ideas might be applied to the development of large computer systems or computational artifacts. The ideas we discuss are still incomplete, and their implications for the design of computer systems still quite speculative. Nonetheless, we believe that even our skeleton, incomplete as it is, provides potential lessons for the design of systems that have a variety of tasks and goals, that must run unattended and autonomously, and that need high reliability. Indeed, consideration of the design constraints on autonomous robots was one of the driving forces that led to this work. (8-13)

The three levels that we propose we refer to as the Reaction level, the Routine level, and the Reflection level (Figure 1). Processing at each level serves two different functions: evaluation of the world and what is happening in it--affect; and the interpretation of what is happening in the world--cognition. Higher levels involve greater depth of processing and concomitant slower processing. As shown in Figure 1, cognitive and affective information flows from level to level. Control information, in the form of activation or inhibition, flows downward.

[FIGURE 1 OMITTED]

The lowest level: Reaction. The Reaction level consists of the lowest-level processes. In animals, these processes are genetically determined and innate. No learning occurs. The Reaction level comprises immediate responses to state information coming from the sensory systems. Its function is rapid reaction to the current state.

The Reaction level monitors the current state of both the organism and the environment through fast, hard-wired detectors that require a minimum of processing. When it detects problematic or dangerous situations, it interrupts ongoing higher-level processing (if there is any), it heightens arousal, and it initiates an immediate response, or response preparation, along with a concomitant diversion of resources.

The output from the Reaction level is a set of fast and relatively simple interrupts, affective signals, and motor actions. Because of the rapid and relatively simple processing, the Reaction level cannot determine causes or do much more than respond in a simple pattern-directed manner. This level is the earliest of evolutionary processes, and in simple animals it is the only processing that occurs. In higher animals and humans, interrupts from the Reaction level trigger higher levels of processing (at the Routine and Reflection levels) in order to determine the cause and select an appropriate response. Responses at the Reaction level can be potentiated or inhibited by inputs from these higher levels, and they can habituate, reducing sensitivity to expected signals.

The mid-level: Routine. In humans, the Routine level is the level of skilled and well-learned, largely "routinized" behaviors. This level is the home of most motor skills, including language generation. The Routine level is quite complex, involving considerable processing to select and guide behavior. It must have access to both working and more permanent memory, as well as evaluative and planning mechanisms. Inputs to the Routine level come from the sensory systems, the Reaction level below, and the Reflection level above in the form of control signals (inhibition and activation). The Routine level can both inhibit and activate Reaction level responses and can pass affective information up to the Reflection level when confronted with discrepancies from norms or routine expectations.

The Routine level performs assessment, resulting in values on three dimensions, which are referred to in the scientific literature on affect and emotion as positive affect, negative affect, and (energetic) arousal. (14) Many emotion researchers now agree that positive and negative affect are essentially independent dimensions (15) as when the motivation of a person on a diet to devour a delicious-looking cookie (a source of positive affect) coexists with the motivation to avoid the same, fattening, cookie (a source of negative affect).

As alluded to above, a key feature of the Routine level is that of default expectations. When these expectations are not met, the system can make adjustments and learn. We return to this point later in our discussion of possible applications. But note the power of expectations in signaling potential difficulties. In humans, these expectations trigger affective processes that play an important role at the higher level of processing.

The highest level: Reflection. Reflection is a metaprocess in which the mind deliberates about itself. That is, it performs operations upon its own internal representations of its experiences, of its physical embodiment (what Damasio (1) calls the "body image"), its current behavior, and the current environment, along with the outputs of planning, reasoning, and problem-solving. This level has input only from lower levels and neither receives direct sensory input nor is capable of direct control of behavior. However, interrupts from lower levels can direct and redirect Reflection-level processing.

There is some evidence that affect changes the processing mode for cognition. The mechanism is neurochemical stimulation that adjusts the weights and thresholds that govern the operating characteristics of the cognitive mechanisms, biasing them and changing the nature of the ongoing processing. These changes influence how higher-level processing takes place, the locus of attention, and the allocation of attentional resources. Thus, negative affect, especially when accompanied by high arousal, appears to lead to more focused and deep processing--depth-first processing. In the extreme case, this type of processing leads to the "tunnel vision" of stress. In contrast, positive affect appears to lead to broad, more widely spread processing--breadth-first processing. As a result, humans have enhanced creativity when in a pleasurable state. (16,17) Both changes are, on average, evolutionarily adaptive (one being consistent with increased vigilance, the other with increased curiosity), even if at times they are counterproductive.

Note that we propose that Reflection has only indirect control (mediated through inhibition and activation) over behavior emanating from the Routine level. The mechanisms of this control have been explored more fully by Norman and Shallice. (180

Implications for machine design

Our artificial systems today have something akin to the three different levels of Reaction, Routine (action), and Reflection, but they do not distinguish between affect (evaluation) and cognition (understanding). In this section we discuss how a model of affect and cognition along the lines of the one we have proposed might apply to machines. Specifically, we suggest that affect can improve overall systems behavior, particularly in complex or difficult environments.

The Reaction level in machines. Reaction is the home of built-in sensors, usually with prewired or preprogrammed, fixed responses. This level is necessary for safety and other critical considerations for which a rapid response is essential. The Reaction level is essential to machine operation, and indeed, is already pretty well recognized and implemented. It is common for computer systems to monitor power and temperature, hardware functioning, and checksums. In robots and other mobile systems, Reaction-level devices include contact sensors and cliff detectors that prevent the devices from hitting other objects or falling down stairs.