Article – Intelligence without representation; Brooks
Brooks persepective is that a problem is solved by not having representations but by
having more basic behaviours.
Bold claims about what’s wrong with AI according to Brooks:
- It focuses on a Toy world rather than the real world
- Presumes a highly abstract sense of the world, and removes it from the actual
world
- Defining an AI problem; once a hard problem is solved, it is no longer defined
as an AI problem
According to Brooks, introspection is inaccurate; if you play chess, by introspection
you can think about it how to solve it, but it’s inaccurate
Alternative is that we should develop systems incrementally, build from the basics,
ensuring that the basic parts are correct, before you add another part.
Layers update each other, each layer has basic abilities to achieve goals, they are
build on top of each other, without a central control (this is different from most AI
systems). There are lots of autonomous bits that are built upon each other.
A layer higher up can inhibit a lower layer, with minimal communication between
layers, but there is communication.
Inspirations in this way of thinking → let’s look at the world; evolution gets us to very
basic organisms, higher level organisms are built upon these. We look at evolution as
a design process; caused billions of years → structure of the structure of the
structure. Maybe we should do the same with AI;
When we find out the basics, higher level problems become easier.
Are representations harmful? Is it the case that Brooks approach could be used to
build a Nigel? → No, to some level, representations are useful; you probably need to
represent it in some way. Some of Nigels design might be inspired by Brooks, some
not
Probabilistic is the opposite of Brooks → you have a probabilistic representation
of the world, with a central control running on probabilistic reasoning.
Brooks persepective is that a problem is solved by not having representations but by
having more basic behaviours.
Bold claims about what’s wrong with AI according to Brooks:
- It focuses on a Toy world rather than the real world
- Presumes a highly abstract sense of the world, and removes it from the actual
world
- Defining an AI problem; once a hard problem is solved, it is no longer defined
as an AI problem
According to Brooks, introspection is inaccurate; if you play chess, by introspection
you can think about it how to solve it, but it’s inaccurate
Alternative is that we should develop systems incrementally, build from the basics,
ensuring that the basic parts are correct, before you add another part.
Layers update each other, each layer has basic abilities to achieve goals, they are
build on top of each other, without a central control (this is different from most AI
systems). There are lots of autonomous bits that are built upon each other.
A layer higher up can inhibit a lower layer, with minimal communication between
layers, but there is communication.
Inspirations in this way of thinking → let’s look at the world; evolution gets us to very
basic organisms, higher level organisms are built upon these. We look at evolution as
a design process; caused billions of years → structure of the structure of the
structure. Maybe we should do the same with AI;
When we find out the basics, higher level problems become easier.
Are representations harmful? Is it the case that Brooks approach could be used to
build a Nigel? → No, to some level, representations are useful; you probably need to
represent it in some way. Some of Nigels design might be inspired by Brooks, some
not
Probabilistic is the opposite of Brooks → you have a probabilistic representation
of the world, with a central control running on probabilistic reasoning.
