Neural Networks

Intelligence as Adaptive Behaviour: An Experiment in Computational Neuroethology is a book written by Randall D. Beer in 1990. It explores a methodology for constructing autonomous agents. Beer's approach to this "lies in the direct use of behavioural and neurobiological ideas from simpler natural animals to construct artificial nervous systems for controlling the behaviour of autonomous agents" (Beer, 1990, p. xvi).

In this case, the chosen simpler natural animal is a cockroach. My first student residence experience made me all too familiar with these little guys. Luckily, the computer cockroach is much cuter, with a simple design inspired by the real thing. The cockroach is programmed using neural networks to wander around its environment and use edge following with its antenna to navigate around obstacles. When hungry, it follows an odor gradient to locate food and ignores it if satiated.

The work in Beer's book was implemented in C to run in DOS by Pat and Greg Williams in 1991. The simulation allows one to set up an environment for the insect to explore. The state of the insect, such as energy/hunger, will change its behaviour based on what its sensing.

The video below shows an example environment for the insect using this simulation.

This is only a simple summary of how the program works, and the book covers much larger territory, including animal neurobiology, development of a heterogeneous neural network, controllers for wandering, recoil, edge-following and feeding, and the problem of behavioural choice.