Brooks, R. A. (1991). "Intelligence Without Reason," Proceedings of 12th Int. Joint Conf. on Artificial Intelligence, Sydney, Australia, August 1991, pp. 569--595.

@InProceedings{brooks:91b,
  author =       "Rodney A. Brooks",
  title =        "Intelligence Without Reason",
  pages =        "569--595",
  ISBN =         "1-55860-160-0",
  editor =       "Ray {Myopoulos, John; Reiter}",
  booktitle =    "Proceedings of the 12th International Joint Conference
                 on Artificial Intelligence",
  address =      "Sydney, Australia",
  month =        aug,
  year =         "1991",
  publisher =    "Morgan Kaufmann",
  url =          "http://www.ai.mit.edu/people/brooks/papers/AIM-1293.ps.Z",
}

Author of the summary: Kory Hopkins, 1996, khopkins@cs.sfu.ca

Cite this paper for:

• THIS FACT
• THAT THEORY, ETC.

This paper provides an excellent background from which many of the current agent models and theories have been developed. It discusses many of the failings of the traditional Artificial Intelligence approach to intelligent agents and proposes a new bottom-up approach to building agents that are able to react in real time in the real world. Brooks' main argument is that research in Artificial Intelligence has been heavily influenced by computer architectures which has caused the field to be led astray. He feels that by examining biological examples of intelligence and using a new bottom-up approach to AI we may be able to make greater advances than the field has been seeing lately.

When AI was first being investigated most programming was done in machine code. A search routine was easy to program and thus became the foundation of most Artificial Intelligence technology. This is an example that Brooks describes where the AI community has adopted a certain technology based on the computing architecture. Today tree searching is still one of the main techniques in many AI reasoning models. Brooks argues that we have become so accustomed to doing things a certain way that we have forgotten that some methods have been adopted due to computer limitations that do not exist any longer. Yet the AI community continues to work with these limitations. The search method has had the illusion of continuous improvement as larger and faster computers have been developed. But this improvement is just due to having a larger search space because of the improved hardware.

Traditional Artificial Intelligence has tried to create intelligent systems in a top-down approach. This has involved attempting to imitate thought and reason to produce intelligence. Brooks refers to this as a Sense Model Plan Act (SMPA) model. The agents first sense their environment through various sensors into the world and then uses this data to build a complete (as complete as possible) model of the world. Then, using this model, the agent produces a plan to achieve its goals and finally acts on the plan. Most of the agents processing time is spent in sensing the environment and building models. Very little of the time is used in planning and acting. The agents (robots) tended to be very slow and were not able to react to changes in the environment very well since the internal models frequently needed updating and were often incorrect.

The SMPA model is significantly different from how biological systems operate. The SMPA is a highly centralised model whereas biological systems are highly decentralised. The SMPA model is made up of many subsections that are each responsible for a specific purpose such as planning or vision. People are not built in this centralised way and it is often impossible to point at one part as the sole source of a given behaviour.

From this background a new style of Artificial Intelligence research has emerged known as behaviour-based agents. Brooks identifies this new model with the following characteristics:

Situatedness: Traditional Artificial Intelligence has involved the manipulation of symbols that represent the real world. Often the models of the agent and the world were merged together and it was difficult to distinguish the two. Brooks argues that rather than attempt to create a model of the world we should use the world as its own model. The agent should use its sensors to continuously monitor the world and be able to react to changes in the environment in a timely manner.

Embodiment: Most attempts at modelling intelligence have been tested in environments simulated within a computer. Few have been successfully tested in the real world. Most of the demonstrations that have been tested have used an extremely simplified and controlled world. Brooks feels that for true intelligence to be displayed the agent must be able to operate in real world conditions. The main reason for embodiment is that the world grounds regress.

Intelligence: Brooks proposes looking at simpler animals as a bottom-up model to building intelligence. It is the interaction with the environment that displays the intelligence of the system. An example Brooks gives is that the complex behaviour of an ant is more a product of its interaction with the world than with its own internal complexity.

Emergence: Intelligence can not be identified to belong to any one part of a system. It is through the interaction of the components of the system that intelligence is able to emerge.

In Brooks' agents he uses some very specific principles:

• there is no central model of the world created and maintained
• there is no central control module
• there is no separation of the system into specific functions. Instead pieces may perform more than one function
• the world can be a good medium for communication between different parts of the agent

This new thinking about Artificial Intelligence has resulted in Brooks' subsumption architecture. It is a series of layers that operate in parallel. The lower layers control more primitive behaviours such as obstacle avoidance and the higher ones are higher level functions like following a walking person. The lower layers are able to temporarily subsume the higher levels through the suppression of inputs to those levels. The agent is able to quickly react to its environment as its sensors are more directly connected to actuators. Also, time is saved since an internal model does not need to be constantly updated and manipulated.

Traditional Artificial Intelligence has attempted to program intelligence into agents. By using this new bottom-up approach to creating agents it is hoped that intelligence will emerge from simpler behaviours.

Summary author's notes:

• none