This paper provides an overview of the existing trends in AI, the influence of technology and ideas on AI, and presents new ideas to break through the existing constraints. The main points discussed in this paper are as below:
1. The author presents arguments from various perspectives to point out that the present (1991) success and methods in Artificial Intelligence (AI) have been influenced and restricted by the available computer architecture and the technological and psychological assumptions (that might be irrelevant today or later) made by the AI community.
2. The top-down approach towards AI, in which the intelligence is understood and emulated as a combination of discrete notions like thought, reasoning, planning, etc., may not be the best approach.
3. The notions governing intelligence (mentioned above) are perceived notions. These traits of intelligence have emerged over the time from simple day-to-day activities of living beings by its interaction and existence in the real world. The author suggests that instead of using the contemporary methods in AI, the use of situated and embodied robots, with emerging intelligence, is more suitable for fulfilling real world expectations.
It should be noted that the current time being referred to in this summary is 1991, the time of this memo. We exclude more recent advances in AI in order to keep the discussion relevant to the time of this paper and the status of AI at that time. Though the first point makes the larger portion of the paper, the paper aims at using the first point and its discussion to derive and conclude to the second and third points. The above points and the main arguments related to them are discussed next.
1. Technological assumptions and constraints
Since early times in AI, it was assumed that if the simple task of operating the robots in static environment is accomplished, the more complicated dynamic environment tasks can also be tackled. Not only a lot of computation and research time was thus wasted in creating environment models (which was far from the real world), the extension from static to dynamic situations was also never accomplished. It has since been realized that the problem of dynamic environment cannot be handled by building internal models of environment. Further, the AI community constraints itself and is unable to emulate the biological intelligence in the representation of the environment model.
It has been conventionally very difficult to model the real world in its actual form. Almost all attempts to model real world are oversimplified into smaller, approximate, and unrealistic models. The author suggest as an example the autonomous land vehicle project by DARPA, where the real 3-D dynamic model of road was compromised to two separate simple components: using 2-D images to identify the road regions and using servo-control motors to remain in those picture-identified road regions. Further, most AI systems use well defined objective internal models, with individually identified and tracked entities, which might not be sufficient as well as dynamic to model the ever-changing, unpredictable and very complicated real world. However, it should be beneficial to use the world itself as a model by making the robots refer to its sensors repetitively to learn (instead of referring to an internal model). Thus, they may have simpler perceptual process and do not need object classification.
In cybernetics, where the attempts were made to understand the behavior of organisms, the approach was to model the whole setup in the form of differential/integral equations and boundary conditions. Though it was understood that the organism/machine and its environment need to be modeled together, the boundary conditions remained largely static, and thus limited its relevance to the real world. Further, the failure of these approaches indicate that intelligence should use different and more abstract tools of analysis.
In the earlier times, besides the problem of representation of the model and intelligence, the major problem was implementation of the robots. In the times of vacuum tubes, implementing stand-alone robots using them was difficult. Later, with advance in technology, the von-Neumann architecture of computers was almost universally adopted for implementing the robots. Unlike the von Neumann architecture, the biological systems have large number of parallel units with very less depth and slow computation, the effects produced by biological systems are results of very few sequential processing steps, and the complete parallel system is dynamic and rapidly changing. Despite this knowledge and the development of parallelization technology, contemporary AI continues to use von Neumann architecture, which has large amount of high speed inactive memory communicating with a high speed processing unit over a very narrow channel.
Turing had long back suggested that a robot should mimic the sensory organs and brain (neural network), but concluded that such emulation would be a herculean task. He then suggested intelligence in terms of symbolism, and proposed minimax approach for decision making. Despite the present sensor advances, many researchers still resort to full-dimension tree-search and minimax/minimization approaches. The original suggestion of Turing was forgotten.
Similarly, in the field of artificial neural network, people advanced from simple feed-forward loops to back-propagated networks. However, despite the propositions like Connection Machine with large local memories, each of one-bit processor, which are highly parallel and fast, AI community still uses complicated, several layered, feed forward, feedback, and back-propagated neural networks.
Another assumption made by the AI community is that the brain consists of simple electrical units which receive input and produce an output. It is often neglected that these electrical circuits and their interactions are not as simple as modeled by the community. These electrical circuits are constantly in a soup of hormones and a mess of various processes occurring simultaneously. On receiving a trigger, the complete soup gets activated and lots of activities take place simultaneously, and the output observed is the behavior that has emerged from a gamut of complicated simultaneous processes. Thus, the model used by AI community to model the brain is vastly different and over-simplistic as compared to the biological model.
2. Top-down vs. Bottom-up approach
Using these various arguments, the author suggests that the AI community has restricted itself from trying something beyond the assumptions, perceived constraints, and traditional practices/ideas in the AI community. The author also brings forth the necessity to use low depth, wider, parallel implementations. Further, he suggests the use of dynamic world model, in which the robots are a part of model, dynamically changing themselves and their environment. The author argues that the top-down approach of having a learning module, a planning module, an analysis module, is not similar to the biological systems. The biological systems comprise of various small simple modules, performing simple tasks requiring simple and few computations, which produce an intelligent behavior by collectively and simultaneously interacting among each other and the environment.
In essence, he suggests that instead of looking at the AI as implementation of different traits of intelligence like reasoning, learning, planning, it is better to implement simpler task-oriented modules that interact among themselves and the environment, such that the desired traits of intelligence emerge from their collective functioning.
3. Intelligence without reasoning
Based on the above (specifically point 2), the author says that intelligence can be achieved without specifically requiring the implementation of reasoning or learning or other traits of intelligence. The author suggests that systems can be built using reactive intelligence, where the systems have to respond to the unpredictable and ever changing world (including themselves) in a robust manner and in a time frame similar to biological systems. He calls such robots as behavior-based robots and proposes four key ideas regarding such robots:
a. Situatedness: Behavior-based robots are situated in the real world and form a part of the real world. Thus, the world influences them and they influence the world. Their situatedness in the real world creates an interdependent dynamic system. The robots do not rely on any internal model of the world. They rather use their sensors continuously to learn about the world and their actuators continuously keep changing the world dependent upon the behavior of the robots.
b. Embodiment: The robots need to directly interact with the world, and thus need to have a complete body of its own with all the possible sensors and actuators. Only fully equipped robots can deal and interact with the real world. Further, the internal symbolism and processing inside the robot are meaningless without the physical interaction with the external world. Since the embodiment requires to deal with all the details and issues of real world, the functional robots already proves the capability to deal with the random perceptual data generated in the real world and is thus realistic.
c. Intelligence: The intelligence is not solely determined by the computational capability. It is also inherently present in the manner in which an organism/robot interacts with the dynamics and uncertainty of the real world events and problems. Intelligence at the smaller organisms’ level does not require reasoning or complex computational capabilities. Rather, at their level, it is defined by the sustainability and adaptivity of the organisms. Thus, it is more reasonable to develop lower level intelligence instead of trying to emulate higher level human-like intelligence directly.
d. Emergence: Intelligence is not a centralized module. There is no central location that can be pinpointed for being responsible for intelligence. There is no reasoning, learning or strategizing module. The brain looks like a huge indecipherable collection of connections and loops. However, there are simple modules for accomplishing small tasks like standing up, avoiding collision, etc. However, intelligence is a perception of the overall behavior of the system which is caused by the complicated interaction of various internal simplistic modules among themselves and with the environment.
Through this paper, the author motivates the AI community to reassess the methods and ideologies used in contemporary AI, and encourages the community to think towards embodied intelligence.