Examining the Validity, Verity, and Relevance of Rodney A. Brooks’s Argument Against the Necessity of Representation in Intelligent Systems

by Zoë N. Finelli, Cognitive Science

The necessity of representation in intelligent systems is a crucial debate within the field of artificial intelligence. This is due to the reasoning and logic-based processing that traditional intelligent systems tend to possess, which is controlled by a central-control system, therefore exemplifying top-down processing in cognition. However, a series of simple, behavior-based systems that does not utilize a central-control system and in turn functions as bottom-up processing in cognition, provides a contrasting approach. Rodney A. Brooks, a well-known roboticist, argues in favor of the latter real-time, reactive approach and against the necessity of representation in intelligent systems. He concludes that a behavior-based intelligence, which he coined as ‘subsumption architecture,’ demonstrates intelligence without the use of representation, thus solving the shortcomings of good old-fashioned AI. This paper analyzes Brooks’s argument, the findings from his research, and the verity and relevance of his work in current artificial intelligence research.

artificial intelligence, artificial general intelligence, subsumption architecture, knowledge representation, intelligent systems, PSI theory

In the paper “Intelligence without Representation,” Rodney A. Brooks challenges the direction of artificial intelligence (AI) research in the late 20th century by arguing against the necessity for representation in the creation of intelligent systems. In his research, he concludes that traditional intelligent systems which rely on internal symbols and models are computationally slow, struggle to handle real, dynamic environments, and assume that intelligence must begin with abstract reasoning. To solve these issues, Brooks proposes a solely behavior-based intelligence, which consists of a collection of simple systems that do not utilize symbolic representation or a central control system. While Brooks’s approach greatly contributed to the shift in AI research from internal representation-based intelligent systems to real-time, reactive intelligent systems, I would like to question the validity and current1 relevance of his work and argue for the necessity of representation within the field of AI and creation of intelligent systems.

First, let us begin by defining a few key concepts. In this paper, ‘artificial intelligence’ or ‘AI’ will be used when describing the field as a whole, while an ‘intelligent system’ will refer to specific technologies, such as software, hardware, robots, sensors, etc. When discussing ‘traditional intelligent systems’ below, we will still be referring to specific technologies, but more importantly those deemed ‘classical’ or ‘good old-fashioned AI,’ which rely on symbolic reasoning and logic. Brooks’s intelligent systems, which he deems as his ‘Creatures,’ are not classified as traditional intelligent systems. Instead, as briefly mentioned above, they utilize a behavior-based intelligence that Brooks coined ‘subsumption architecture,’ which we will discuss in further detail later. The term ‘representation’ will be used when discussing how (e.g., symbolic, graphs, object & relationship), or even if, certain intelligent systems store information. When discussing unspecified representation in relation to Brooks’s work, we will most often be referring to ‘symbolic representation,’ which stores knowledge as symbols and makes decisions based on logical rules.2

Now, let us also define the concepts of ‘necessity’ and ‘sufficiency’ according to the logical operator – the material conditional3, which may be used in forms such as ‘sufficient condition,’ ‘necessary,’ etc. I would like to define necessity as the undeniable dependance of the validity or existence of an entity (#1) on another entity (#2). Therefore, entity #1 can not exist under any circumstance without the existence of entity #2, as we have seen above with our antecedent and consequent. Furthermore, I would like to define ‘sufficiency’ as an entity (#1) adequately fulfilling a relational role to entity (#2), but not being the sole, indispensable cause for the existence of entity #2. Thus, entity #1 may exist or may not exist, but the existence of entity #2 is not affected by the existence of entity #1. Therefore, I would like to acknowledge that Brooks does not argue that representation is a non-necessity in the creation of artificial intelligence, but rather it is a sufficient condition in the creation of artificial intelligence, as he never explicitly states that traditional intelligent systems are not artificial intelligence, but rather that they are not as intelligent or successful as his own intelligent systems, which I will expound upon in the future. In other words, Brooks is allowing for the possibility of non-representational intelligent systems, which we will now discuss.

Traditional intelligent systems use neural webs and networks to store information and some sort of central system, which acts as the main controller and decision-maker of its set of systems. While these systems do not actually learn or adapt very well to new or ambiguous situations, they still implement reasoning and problem-solving to logically draw conclusions. This leads to behavior which is rule-based, explainable, and interpretable. However, Brooks’s systems are composed of finite state machines (FSMs), which can only possess one state at a time, represented by a number and out of a limited number of states, and sensors that directly receive information from the physical world that is not stored. This means that the behavior of these systems is based solely on sensory inputs, and behavioral transitions only take place when certain events occur, which trigger the change from one state to another. The union of these systems forms ‘layers,’ representing a bottom-up model4 of intelligence where information from the environment is received through the sensors, creating a constant feedback loop between the environment and the intelligent system. Therefore, “there are no variables that need instantiation in reasoning processes. There are no rules that need to be selected through pattern matching. There are no choices to be made. [And] to a large extent, the state of the world determines the action of the Creature” (Brooks, 1991, p. 406). This process, better known as subsumption architecture (SA), is Brooks’s solution to the shortcomings of traditional intelligent systems.

To properly analyze Brooks’s solutions, we will now introduce two of Brooks’s Creatures, ‘Allen’ and ‘Herbert,’ to demonstrate the performance of SA. As Brooks details in his reports, both Allen and Herbert navigate office space without the need to store any information, and therefore do not have representation, nor have a central system. Herbert is also able to accomplish the added task of picking up soda cans and grabbing cans when they are placed in his hand. This exemplifies the use of SA because the Creatures interact with their environment on an extremely localized level, and they rely on sensory inputs around them (e.g., the office space, a can given to Herbert) in order to transition between states.

The Creatures also do not store any reached states (i.e., utilizing representation) and only change states based upon their predefined rules of interaction (e.g., stop, wander, reverse). These outward behaviors may be perceived by humans in the office space as ‘learning’ the layout of the office and how to complete tasks, although inwardly the Creatures have no stored memory or three-dimensional representation of the space. This leads Brooks to conclude “that there need be no explicit representation of either the world or the intentions of the system to generate intelligent behaviors” because the behaviors are only determined by the perception of the human (Brooks, 1991, p. 406). I will now question the validity of Brooks’s conclusion by raising several counters and responses to potential rebuttals of my counters.

To begin, I would like to acknowledge that Brooks is clear from the start of his paper that he has “no particular interest in demonstrating how human beings work” although still referring to them as “successful autonomous agents,” as he criticizes the focus of artificial intelligence being on the recreation of human intelligence (Brooks, 1991, p. 401). Brooks writes that he wishes “to build completely autonomous mobile agents that co-exist in the world with humans, and are seen by those humans as intelligent beings in their own right,” which marks the birth of Brooks’s Creatures (Brooks, 1991, p. 401). In response to this statement, I believe it is necessary to question Brooks’s motivation behind his research.

What is the goal of the field of artificial intelligence? To recreate human intelligence, to simply imitate the intelligence of another high-functioning animal, or perhaps something else? Yet, Brooks writes of the biological evolution of simple human actions, arguing “that mobility, acute vision and the ability to carry out survival related tasks in a dynamic environment provide a necessary basis for the development of true intelligence” (Brooks, 1991, p. 397). So, while Brooks may be using humans to establish his requirements for intelligent systems, he is still not concerned with the replication of human-level intelligence. This is particularly concerning because without the proper establishment of the definition of ‘intelligence’ within his work, Brooks’s argument loses validity due to the unclear establishment of the goal of his research.

If we consider Norbert Wiener’s metaphor5 of the processes of the mind being an unknowable ‘black box’ when analyzing Brooks’s intelligent system, his findings may have resulted in a less contradictory conclusion (N & Wiener, 1949). This is because he would have created an intelligent system in favor of philosophical behaviorists of the time and their definition of intelligence. However, I, along with the current direction of research in AI6, do not define intelligence solely based on behaviorist ideals. Instead, I would like to view Brooks’s conclusion through a functionalist7 lens and examine how Brooks concludes that his Creatures are successful in demonstrating intelligence without representation.

We have seen above that Brooks refers to his Creatures as producing intelligent behaviors, but he fails to blatantly state that he believes his Creatures are simply intelligent or possessing of intelligence. However, he still attempts to solidify his Creatures’ intelligence by taking the behaviorist position that “intelligence can only be determined by the total behavior of the system and how that behavior appears in relation to the environment,” and that “intelligence is in the eye of the observer” (Brooks, 1991, p. 419). One may argue that intelligent beings (e.g., humans) produce intelligent actions and behaviors. However, does this also mean that an entity that produces intelligent actions and behavior should be considered intelligent?

I believe that the very obvious answer to this question is no, therefore disproving Brooks’s conclusion of the non-necessity of representation. For example, say there is a programmed thermostat in a house responding to dynamic temperatures during a twenty-four-hour period, keeping the temperature between a specified range by turning the heat and air-conditioning on or off. It is not storing any record of the temperatures yet constantly adapting to the sensory input of the environment surrounding it, which may be deemed by some as intelligent action. So, by Brooks’s criteria, this entity is demonstrating intelligent action without representation, thus is also intelligent. However, the thermostat is not learning, reasoning, thinking, etc.; it is only responding to sensory input through pre-programmed states and performing actions in response, therefore not demonstrating intelligence when analyzing this example through a functionalist lens. And, to any well-functioning human, the argument of a programmable thermostat being an intelligent entity8 sounds utterly ridiculous because of the way in which we have described its functioning.

After our example above, I believe it is important to also establish a definition for ‘learning.’ A counter of relevance may include the question: If Brooks states that intelligence is determined by the observer, does this mean that learning is also determined by the observer? For this paper, I would like to note that some elements of philosophy are inherently subjective due to the questions that they address (e.g., philosophy of mind). If we leave defining the concept of ‘learning’ to the observer, one may conclude that the thermostat is intelligent, based on opinion alone. However, to avoid an endless philosophical spiral, I would like to refer to learning as ‘the ability for intelligent systems to adapt internal representations and functions in accordance with new experience and knowledge in order to improve performance and/or behavior.’ Now, let’s put the intelligence of one of Brooks’s Creatures to the test in a new example below.

Say that I created a Creature that relied solely on subsumption architecture to complete the task of traveling through a maze. However, this maze was not just a static maze, similar to the office which Brooks’s Creatures navigate. This maze has a trap door in the middle of a pathway that opens on a three second timer. If we were to put a Creature in this maze, the only state that its FSM would be able to interpret is the current state of the door, which in this case we will assume could be ‘open door’ or ‘closed door.’ This is due to FSMs not storing information as a representation and only being able to hold one state at a time, as discussed earlier. Since the Creature possesses no internal representation, it also lacks memory of encountering, if ever, these states before. Even if the Creature is constantly receiving sensory input of the state of the door, it will never be able to learn, nor even recognize, the timer pattern of the trap door opening and closing and navigate it with accuracy. Therefore, the Creature will only be able to perceive the state of the door as it is approaching it and will always rely on the chance that the door will not open. If the Creature does happen to perceive the door as closed and drives forward onto the door, which now suddenly opens due to the timer, the human watching the Creature fall through the trap door to its death will most-likely not deem this as intelligent action either. 

This demonstrates that the Creature has no way of predicting the opening of the door; it can only perceive the states of ‘closed door:’ safe, move forward and ‘open door:’ danger, stop. However, if the Creature was simply able to store encountered states (i.e., internal representation) and properly utilize a simple algorithm for detecting the timer on which the trap door was set, it would now be able to observe and store the timer pattern for future navigation. Thus, here we see the explicit necessity for representation in intelligent systems.

A rebuttal to this example may mention that this is still a static environment in which intelligent systems have traditionally been tested because the location of the maze is fixed, and the only dynamic movement is an obstacle. However, I would like to point out that Brooks’s main argument about static environments is that traditional intelligent systems are not able to succeed in the ‘real world,’ which he deems is a more important environment and of greater difficulty to succeed in. But he does not state that his Creatures would not be able to succeed in, for example, the maze. If Brooks would like to argue that intelligence is in the eye of the observer and “determined by the dynamics of interaction with the world,” I would like to respond that such a simple example where his Creatures do not succeed is indeed a world in which proves his Creatures are not intelligent (Brooks, 1991, p. 418).

Finally, I would like to give Brooks recognition for his work because I do believe that it is still relevant in current AI research, but that he fails to establish a valid and convincing argument for the correct role of his subsumption systems. As we discussed above, intelligence is often9 associated with learning, in humans and in other animals. We have concluded that Brooks’s Creatures are not learning about their environment, but rather perceiving their environment, which in turn produces intelligent action that makes them seem as if they are indeed learning. The FSMs used in Brooks’s Creatures rely on their pre-programmed states, which leads to a constant feedback loop of sensory input. While Brooks attempted to argue this as demonstrating intelligence, I would like to deem this process rather as a replication of biological instinct.

As humans, we are instinctually experiencing a constant, subconscious feedback loop of information that our body is adapting to. For example, if a stove is hot, we are biologically programmed to remove our hand. Of course, one could say that we are making a choice to remove our hand due to pain being an emotion and the choice of physical movement, and that the biologically programmed pain receptors and most common desire to get away from the hot surface is a survival instinct. And while this action is intelligent in a sense that evolution is fascinating and we have evolved to feel pain, I would not deem this action to remove my hand as intelligent, but rather as the biological instinct that intelligently evolved.

Now, this raises some particularly interesting points. Is there a place for biological instinct in current AI research? I would say that the short answer is yes, but the long answer is no, not in all facets of AI. Subsumption architecture has mostly been utilized in robotics to achieve real-time environment responsiveness. Beyond this, however, the current applications of SA are dwindling due to the success of neural networks, hybrid logic-based systems, deep learning, reinforcement learning, natural language processing, etc. And while these approaches are making extraordinary advancements in specific implementations of AI, they are all still falling short in one area of application: artificial general intelligence (AGI), which refers to an intelligent system that possesses human-level intelligence. I believe this is because these architectures lack the integration of biological human instincts yet are still trying to replicate human cognition.

Let us now briefly discuss three of the leading AGI architectures to analyze the direction of the field of AGI. SOAR architecture, created in the 1980’s by John Laird, Allen Newell, and Paul Rosenbloom, is a symbolic cognitive architecture that focuses on general problem-solving and decision-making (Laird, 2008, pg. 3–4). Similarly, ACT-R (Adaptive Control of Thought-Rational), introduced by John Robert Anderson and Christian Lebiere, is a hybrid cognitive architecture that emphasizes memory, perception, and learning through symbolic and sub-symbolic10 approaches (Carnegie Mellon University, n.d.). Although SOAR relies heavily on symbolic methods to achieve goals and ACT-R incorporates more neural networking, both architectures follow a similar approach, attempting to recreate human cognition through logic and reasoning, the use of a short- and long-term memory, and motivational drives. And while these architectures are widely supported and established, they both have shortcomings, such as over-complexity as more rules are added, rigidity in symbolic reasoning, and limits in real-word applicability.

However, MicroPsi, invented by Joshua Bach, takes a unique path away from these architectures, following the PSI (Personality Systems Interaction) theory11 of Dietrich Dörner. MicroPsi specifically focuses on motivation and emotion, as well as the more traditional components of representation, logic, and reasoning processes (Bach, 2003, pg. 1). As Dörner and Güss discussed in “PSI: a computational architecture of cognition, motivation, and emotion,” current architectures, including SOAR and ACT-R, prioritize either “cognition or emotion or motivation,” but do not implement at all three processes concurrently (Dörner and Güss, 2013, pg. 1). I find this particularly interesting because of the significant portion, if not majority, of human cognition that is shaped by emotional processes. According to the PSI theory, emotions are not only a product of someone’s experiences and perceptions of the world but are largely properties of cognitive modulation: how the brain modulates perception, behavior and cognitive processing (Dörner & Güss, 2013). As a result, the pre-programmed, biological response to a perception and the perception itself is incredibly influential on how humans learn, organize experiences, and simply move throughout the world (i.e., a system’s intelligent action).

Now, it is important to highlight that Bach’s hybrid cognitive architecture and Brooks’s subsumption architecture share two key similarities:

  1. They step away from the traditional AI approach of strictly symbolic representation and logical reasoning.
  2. They emphasize recreating the biological, instinctual processes of human cognition.

And while I do not find Brooks’s argument to be entirely valid, I do believe that Brooks was correct in calling for more focus on instinctual, biological action in the development of intelligent systems, as seen in MicroPsi. Therefore, I would like to conclude by proposing a hybrid cognitive architecture with the integration of SA to represent low-level emotions and motivational drive.

Although SA independently may not be the solution to creating intelligent systems, its integration within an architecture that also utilizes representation may provide a unique step ahead of those currently leading AGI, such as SOAR and ACT-R. Creating a system that employs SA to replicate low-level human emotion would successfully apply the effects of emotion on perception, and in turn intelligent action. Then, methods of representation, such as neural networking, would build upon the immediate reactions of the system produced by the layers of SA, logically organizing and storing experiences and information. So, the system would produce intelligent action based on not only the content of its external experiences, but also from its internal replicated emotional response, serving as its motivational drive and deeper sense of direction. Not only could this provide better handling in real-world application due to less reliance on symbolic representation for direction, but the system would also possess more contextual awareness from the added layer of experiential connection contributed by emotion, leading to stronger problem-solving, better learning, and true emotional awareness.

With this in mind, I do believe that Brooks was correct in calling for more focus on instinctual, biological action in the development of intelligent systems, especially in the field of AGI. However, I do not find Brooks’s argument against the necessity for representation in intelligent systems to be accurate or valid, nor do I support it. While subsumption architecture may not independently be the solution to creating intelligent systems, I believe it will be an integral part of reproducing true, human-level intelligence in regard to emotion and instinctual action.

Notes

  1. I would like to specify that ‘current’ here is referring to 21st century artificial intelligence research, occurring near the writing of this paper. ↩︎
  2. The conditional, also known as the ‘if, then’ statement, can be described as a horseshoe-like symbol that represents the relationship between two entities: the antecedent and the consequent. When assigning the truth-values, true or false, to the antecedent and the consequent of the conditional, any combination of truth-values on any truth value assignment will result in a truth-functionally true sentence, except for the case in which the antecedent is true and the consequent is false. Characteristically, this means that the antecedent can only be true when the consequent is true. However, the consequent may be true when the antecedent is either true or false. ↩︎
  3. This is due to symbolic representation (what is most often associated with traditional intelligent systems) being what Brooks specifically rejects. ↩︎
  4. A bottom-up approach refers to a model of intelligence that emerges from sensory input and experience, obtained through interactions between simple units and which utilizes self-organization rather than pre-defined knowledge (e.g., subsumption architecture, neural networks). On the other hand, a top-down approach begins with structured knowledge, rules, and logic, and then applies reasoning to reach a conclusion (e.g., symbolic AI, good old-fashioned AI). ↩︎
  5. This metaphor was first used in the field of psychology in 1961 to describe the mind as the unknowable process between input and output (i.e., stimulus and response). It was then employed by behaviorists, such as B.F. Skinner and John Watson, to argue that psychologists should only be concerned with observable behaviors, rather than attempting to understand internal mental states. This view of the approaching the mind is known as ‘behaviorism.’ ↩︎
  6. While it is plausible to cite several sources here explicitly showing the ‘current’ direction of AI, I would like to leave this statement open for interpretation and personal enlightening. ↩︎
  7. Functionalism refers to defining intelligence based on what a system does and not what a system is made of (e.g., a computer does not possess biological brain structures such as neurons, but utilizes neural networks in order to learn, think, reason, or perform other intelligent-like actions). ↩︎
  8. Although, I would like to note that programmable thermostats are wildly fascinating for a college student trying to save money on her gas bill. ↩︎
  9. Again, I would like to note that this is taking on the functionalist approach to current research in AI. ↩︎
  10. Sub-symbolic approaches refer to those such as neural networks, probabilistic reasoning, or reinforcement learning, which rely on adaptation, learning, and perception-based processing. ↩︎
  11. The PSI theory can be best described as “a formalized computational architecture of human psychological processes” that models cognitive, motivational, and emotional actions and their processes (Dörner & Güss, 2013). ↩︎

References

Bach, J. (2003). The MicroPsi agent architecture. Fifth International Conference on Cognitive Modeling. http://cognitive-ai.com/publications/assets/MicroPsiArchitectureICCM03.pdf

Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47(1–3), 139–159. https://doi.org/10.1016/0004-3702(91)90053-m

Carnegie Mellon University. (n.d.). ACT-R. http://act-r.psy.cmu.edu/

Dörner, D., & Güss, C. D. (2013). PSI: A computational architecture of cognition, motivation, and emotion. Review of General Psychology, 17(3), 297–317. https://doi.org/10.1037/a0032947

Laird, J. E. (2007). Extending the Soar cognitive architecture. https://doi.org/10.21236/ada473738

N, E., & Wiener, N. (1949). Cybernetics, or control and communication in the animal and the machine. The Journal of Philosophy, 46(22), 736. https://doi.org/10.2307/2020260

Citation Style: APA7

Back to 13.1 Table of Contents