OpenAI just released a new research paper in which agents develop their own language
High level Description
- AI agents are taught to create a language in a set of simple worlds and are rewarded for achieving goals that are best completed by effectively communicating with other agents...i.e. the end result is that they develop a shared language
- The experiment is represented as a cooperative — rather than competitive — multi-agent reinforcement learning problem. At each time step, the RL agents can take two kinds of actions — environment actions, like moving around or looking at things, and communication actions, like broadcasting a word to all other agents.
- Issues that the agents faced:
- They would create a single utterance and intersperse it with spaces to create meaning. This Morse code language was hard to decipher and non-compositional.
- They would try to use single words to encode the meaning of entire sentences.
- They would invent landmark references not based on color, but other cues such as spatial relationships. This behavior could cause problems if the geography of the worlds the agents lived in was changed