Project: Emergent Communication in Social Games 

Key information:

Abstract: 

With the world becoming more interconnected, we can see the rise in the number of real-world applications that can be managed by multiagent systems. In our future vision, the intelligent agents mentioned here could be deployed onto specific platforms, like asteroid mining drones. The agents utilised there would have to converse together and agree on actions that they would take to maximize the scientific and economic return of such missions. They would also need to be adaptive and resilient to any changes in the environment, or unforeseen objects. For both of these objectives, effective communication is key. Furthermore, as they would be operating far away from any energy source or possibility of maintenance, they would need to be energy efficient.

All agents, deployed in any multiagent environment, need to cooperate, compete, or both, to achieve the best results. To do either well, they have to be able to communicate effectively. However, when we try to impose any language onto those agents we run into issues of how much information to pass on, and what words or protocols to include. Moreover, the languages that we do impose may be incomplete or outdated by the time we can deploy the agents.

To solve these problems, we will be investigating what is called emergent communication. It is the extreme case of allowing intelligent agents to develop their own communication protocols and languages. Instead of providing the agents with any basis for their language, we allow them to create everything from scratch, as required by their environment. This allows them to fully match the needs of what they need to communicate about in their surroundings, while also increasing the efficiency of their communication. Through this, the agents will be able to interact better with each other, and increase their efficiency not only in the communication aspect but also in their performance overall. To research this, we will be using highly complex environments to train our agents, which will require them to develop languages with interesting properties.