standard IQ tests, though possessing several desirable properties for measuring the intelligence of humans, are not suitable for measuring machine intelligence
some might suggest developing entity-specific tests for testing intelligence, but this does not work well for testing entities which are very different from others. also, our concept of intelligence is rapidly evolving
how can we develop a concept of intelligence that is applicable to all kinds of systems?
from comparing several definitions of intelligence, they arrive at an informal working definition of intelligence: _intelligence measures an agent's ability to achieve goals in a wide range of environments_
they then define a framework for intelligent agents. it consists of an agent, which takes actions, and an environment, which signals the agent with observations. they then add a reward feedback mechanism from the environment, and define goal fulfilment as maximizing the amount of reward
they then define the agent's actions as being of members of a finite set, with the agent's perceptions also being of that set. in other words, they are discrete signals. they also define the reward space as an utility between 0 and 1, with the observation space being a pair: an observation and a reward
the paper then leads into a discussion which makes it clear that the expected reward of the agent which should be used for reckoning intelligence—not the reward actually received. the author defines a "standard way" of formalizing. he then redefines the model in such a way so that the such calculations are not necessary, considered beyond the scope of intelligence
they discuss some of the theoretical problems with the model, such as the possiblity that the agent can take over it and give itself maximum reward (using pain medication as an [weak] analogy)
the [so far, fairly uninteresting] paper suddenly takes a turn for the better when the author conjectures that simple environments are more likely than complex environments, citing ockham's razor. he states that the complexity of an environment can be measured by the kolmogorov complexity of the turing machine which represents the environment. he points out that kolmogorov complexity can be used to create a probability distribution, the algorithmic probability distribution, which he seems to believe is the prior distribution from which we expect our environments to be drawn. this leads to a simple but elegant formula which defines the intelligence of an entity as the sum of expected reward across all possible environments, multiplied by the its probability density in the algorithmic distribution
he then points out that a random agent would have very low intelligence, a very specialized agent would have high rewards in those environments which it is specialized for and low rewards in other environments, and so forth
he then lays out several claims for desirable properties of this measure
one questions the justification for selecting the algorithmic distribution as the prior distribution for environments. furthermore, it is not entirely clear how the lessons of this paper are to be extended to continuous environments. nonetheless, this was a very interesting paper
