A UK scientist has persuaded robots to sing to one another, in research that has implications for co-operating autonomous machines.
"At the end of the day, this is not only about music. Robots could agree collective strategies to perform a task together, and develop learning strategies to come to agreed tasks quickly," Professor Eduardo Miranda of the University of Plymouth told Electronics Weekly.
According to Miranda, the human brain processes different aspects of music in various places, some of which may be specific to music and some of which are not so specific. Functioning altogether, the brain uses these different brain functions to relate heard sequences to remembered sequences.
Miranda's robots mimic human vocal tract and hearing, and are programmed to behave like certain brain functions thought to exist in humans that recognise when a sound they hear is similar to a sound their host has produced.
Initially, the Plymouth robots babble melodies randomly.
When incoming sounds are judged similar to remembered sounds, the robot can respond with an imitation of the incoming sound.
"The task is to produce two sounds alike," said Miranda. "The robots have a set of criteria to select sounds. Certain sounds are discarded and certain sounds are kept."
Up to 20 virtual robots have been run in an experiment, picked randomly in pairs to spend some time interacting in the actual robots - DRK8000s from Dr Robot.
With more cash there would be more real robots, but why not model the whole thing in a computer?
"I have done so," said Miranda who is Professor of Computer Music. "I moved to real robots because I want eventually to have robots that interact with humans. Eventually they will learn from me and I will learn from them."
According to Miranda, attempts so far to persuade machines to make music together have used abstract algorithms embodying pattern generation features suitable for making music: cellular automata and particle swarms for example or music knowledge-based techniques: symbolic machine learning or neural networks, armed with music theory and learning from existing music.
The former tend to produce complex material few people recognise as music, and the latter generally produce music similar to the training pieces.
His approach differs in avoiding manually programming prescribed rules for generating music, but attempts to programme the robots with the ability to develop suitable musical rules themselves.
To keep things simple, only short tone sequences have been allowed in Miranda's experiments, stored as parameters that describe lung pressure, throat position and vocal chord tension for each tone.
In a typical interaction giving pairs 20s to communicate, a group of five robots developed on average 19 tunes each, mostly settling by 2,000 interactions.