a summary of the article “intelligence without representation” by rodney a. brooks (1987)...

13
A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

Post on 19-Dec-2015

226 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

A Summary of the Article “Intelligence Without Representation”

by Rodney A. Brooks (1987)

Presented by Dain Finn

Page 2: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

Initial goals of the AI field were ambitions – completely replicating human intelligence

Gradual realization of what a huge task this is

Focus shifts to specialized sub-problems:Language processingVisionKnowledge representation… etc …

Page 3: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

Brooks (and others) believe human intelligence is too complex to try to decompose with our present knowledge

In favor of a different approach:

Incrementally build up capabilities – each step is a complete (yet simple) system in itself

Each step must be tested thoroughly – in the real world

Brooks believes:

Attempting to come up with models and representations of the real world is the wrong approach

Page 4: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

3.5 billion years ago

2.5 billion years ago

apesprim

atesm

amm

alsdinosaursreptilesinsects

fish and vertebrates

single-celled life

present day

photosynthesis

* Not to scale Expert know

ledgeW

ritingA

gricultureH

umans

550 million years ago

2.5 million years ago

Brooks’ conclusion:

Complex behavior, knowledge, and reason are all relatively simple once the basics of survival - moving around, sensing the environment, and maintaining life - are acquired.

Page 5: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

Brooks: AI researchers tend to factor out motor skills and perception; he believes that these _are_ the difficult problems that need to be solved.

The blocks world: popular AI research concept in the 60’s and 70’s. Everything in the “world” was built from blocks. Criticized for its oversimplification and not being relevant to real-world problems.

New focus on representation: coming up with systems to abstract the semantics of a world, to reduce it to a simple set of pertinent facts that the computer could then work with.

Brooks: it’s the act of creating this type of abstraction that is the essence of intelligence. As long as we do all of the abstraction for our programs, we’re basically still working in the blocks world.

Page 6: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

Brooks’ goal: To build “Creatures” - autonomous mobile agents that exist in the real world. More concerned with an engineering methodology for achieving this than with discovering how the human mind works or with philosophical topics.

Requirements of a creature:

Must cope with dynamic environment

Gradual changes in the environment should result in a gradual change of the creature’s behavior, not total collapse

Creature should maintain multiple goals. Depending on circumstances it will re-prioritize these.

The creature should have some purpose.

Page 7: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

Engineering approach #1 - decomposition by function:

The traditional view is to have one central information processing system with input and output modules connected to it.

Central System

Input module #1

Input module #2

Output module #1

Output module #2

Page 8: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

Engineering approach #2 - decomposition by activity:

Brooks’ idea is to make no distinction between different systems like “vision” and “central systems” – but rather to have individual layers that each produce one activity.

Activity layer 1

Activity layer 2

Activity layer 3

Page 9: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

•Representation is removed with the idea of multiple layers: the lines are blurred between where input and output is occurring. Sensor data is acted on by multiple systems working independently in parallel.

•Low level activities allow fast reactions to dangerous circumstances without any delay of data processing.

•By removing a central control point, there’s less chance of total collapse. Each layer extracts only aspects of the world that it finds relevant. Changes may produce difficulty in a single layer without the Creature completely breaking down.

•Each layer serves its own goal. Each layer continuously monitors the environment and adjusts as needed. Separate hardware for each layer – so adding new goals won’t slow down the Creature’s processing.

•The “purpose” of the Creature is a property of the design of the higher-level layers. There’s no central system selecting from a list of possible goals.

•The Creature is a collection of competing behaviors. It is only the observer who sees the Creature to have a central goal: a pattern of behavior.

Page 10: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

•Creatures must be tested in the real world. A simplified world won’t do.

•Simplified creatures first, then add sophistication.

•With a thoroughly debugged system, add another layer. If there’s problems, it can only come from:

•Existing layer

•The new layer

•An interaction between the new and existing

Page 11: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

•Brooks and colleagues built an initial set of robots to demonstrate their ideas. They use what they call the “subsumption architecture”

•Each layer in the subsumption architecture is a network of simple finite state machines, which run asynchronously and send messages between each other. The arrival of messages or the expiration of timers trigger state transitions.

•Layers are combined using “suppression” and “inhibition”: injecting messages into lower layers and telling them to temporarily suspend normal operation.

•Example of one of his robots:

•Lowest layer: robot avoids hitting physical objects

•Middle layer: robot randomly wanders

•Top layer: robot seeks out distant reachable points

Page 12: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

•This “subsumption architecture” may resemble other approaches but Brooks believes it is distinct:

•Connectionism: an approach that focuses on using multiple processors networked together; the focus tends to be on the connections themselves and the hope of finding a distributed representation in the system.

•Neural networks: parent discipline of connectionism

•Production rules: rule is selected from a rule database using a set of preconditions (variables)

•Blackboard control system: localized knowledge sources put their information on a central “blackboard,” and different sub-units search this repository of information to make decisions.

Page 13: A Summary of the Article “Intelligence Without Representation” by Rodney A. Brooks (1987) Presented by Dain Finn

•Brooks: although promising, many questions remain about this approach:

•How many layers before the interactions become too complex?

•How complex can an individual layer be without using central representations?

•Can the simple finite state machine model support high level functions like learning?