Prediction and Intelligence by David Gamez Intelligence is not a precisely defined property, like mass or charge, and it cannot be directly measured People recognize intelligent behaviour and rank people according to their intelligence, but we cannot point to the intelligence in a brain and we cannot program general intelligence into a machine. To address this issue people often measure intelligence using batteries of behavioural tests based on verbal reasoning, spatial manipulation and mathematics. The results from these tests are typically converted into values of intelligence quotient (IQ) or g-score. The measurement of intelligence through batteries of tests has some plausibility with humans, since we generally agree about which behaviours are linked to intelligence. However, it becomes much more problematic when we want to compare the intelligence of different species. Animals cannot take human intelligence tests, so there has been some work on the development of cognitive test batteries for animals. It might be possible to come up with a plausible set of tests that could be applied to similar animals, but this approach is likely to neglect the different types of intelligence that animals develop to survive in their ecological niche. A measure of intelligence designed for sheep or fish, for example, cannot easily be transferred to birds or bees. These problems become more acute when we attempt to measure intelligence in machines and try to compare the intelligence of natural and artificial systems. A computer that was programmed to outperform humans on IQ tests could be completely incapable of performing any other task that we consider to be intelligent. It is likely to be impossible to use batteries of behavioural tests to compare human and machine intelligence. One approach to this problem is to give up on the idea of a meaningful set of tests to measure intelligence across all animals and possible machines. Instead, we can take humans as our benchmark and rank animals and machines according to the extent to which they match or exceed human intelligence. This is a form of Turing testing. A different response to this problem is to develop universal intelligence tests that enable humans to be compared with other species and artificial systems. In previous work, universal measures of intelligence have been developed based on the ability of agents to achieve goals in different environments. The total intelligence of an agent is the sum of the rewards that it achieves across all possible environments with adjustment made for the complexity of environments. In this talk I will outline the work that I have been doing on a measure of intelligence based on a system's ability to make predictions. Many behaviours linked to intelligence, such as spatial, mathematical and verbal reasoning, require prediction, and our ability to succeed in a variety of environments is closely tied to our ability to predict the consequences of our actions in different environments. A predictive approach to intelligence also fits in well with the recent surge of interest in the predictive brain hypothesis: If brains are intelligent and the brain's core function is prediction, then brains that are better predictors will be more intelligent. The recent successes of artificial intelligence have also been largely based on the ability of machine learning algorithms to generate predictions. Predictive ability can be measured through external tests and I am developing an algorithm that will measure a system's predictive intelligence from its internal states.