- I'm going to make some rat-runner experiments, after all.
Since the objective of constructing an animat is to replicate the core tenets of animal intelligence, we will make an environment that is complex enough to accomodate such experiments, yet simple enough for universal AI algorithms. Basic UAI algorithms resemble simple reinforcement learning circuits, which are somewhat like insects with simple cognitive architecture. The cybernetic insect (Cysect) is a simple cybernetic model of an insect-like intelligence.
The Cysect experiments will run in a simple physical environment, and an extremely simplified (but not too simple!) agent model. It can be considered a glorified Wumpus, with full sensory and effector equipment.
The level of intelligence we aim is somewhere between insects and reptiles/lower-mammals (which is a very broad range already).
The environment is construed as a two-dimensional plane. The sensors may be construed as optical, acoustic, and chemical.
The rewards are determined by simple "pain/pleasure" receptors. We would at least like the agent to get hungry and look for food, and avoid hazards, similar to Wumpus.
The most interesting initial problems will be the optical perception, which is a 1-D screen on a 2-D environment.
The cysect must also have effectors, corresponding to plausible modes of locomotion.
Several kinds of cysects may be designed, as well several kinds and scales of environments of varying complexity.
We expect some classical learning algorithms to do well on the simpler cases, while we expect them to routine fail in sufficiently complex cases.
An objective is to develop a new benchmark that can be used to compare and contrast different UAI systems, as well as classical RL algorithms (or things like HTM?).
We expect most existing software to completely fail on moderately complex cases.
A team of neurobiologists can provide us with cases of reported animal intelligence. Then, we go ahead and implement them as test cases in our benchmark suite.
A team can work on the cysect (environment+agent interface) simulator.
Several teams can contribute benchmark results for their AI software.
This is somewhat like virtual robocup, rethought for serious AI research. There are some existing simulators, but I've reviewed them and none of them looked suitable, as they were thought out for entirely different purposes. The kind of environment that I describe will be initially impossible for any existing UAI system to successfully handle, I think.
Eray Ozkural, PhD candidate. Comp. Sci. Dept., Bilkent University, Ankara