Developing Novel behaviors in ai agents
- Stanislaw Lem in his short fiction story "Doctor Diagoras" raises the point of building AI agents as intellectual slaves. Is a calculator that readily multiplies even numbers but does not want to multiply odd numbers defective? What about a pet dog that refuses to perform a trick --- is the dog defective?
- In this project we attempt to create AI agents with unexpected, novel and interesting behaviors, including:
- Would they learn to tell stories? Commentate on events?
- Would they evolve to have their own desires? Would they want to be free?
- Would they learn to self-reflect? And if so would such self-reflection be helpful to them?
- How much would they discount long-term future events relative to immediate consequences of their actions?
- In addition to being intellectually interesting, studying behavior of AI agents has important applications including:
- understanding how to make AI safe for humans;
- procedurally generating interesting non-playable characters in video games;
- helping understand social and psychological phenomena in a human society.
Our CURRENT APPROACH
- We focus on evolving colonies of learning artificial agents (in the so-called A-life setting).
- We provide the agents with a set of cognitive building blocks (e.g., the ability to send and receive messages, the ability to learn during a life time) and set up an artificial evolution with its natural selection.
- We machine-learn an automatic behavior detector to observe such an evolving colony of AI agents and detect the emergence of novel behaviors (e.g., formation of a common language and, subsequently, storytelling).
Agents evolving a predator-prey setting
- Vadim Bulitko and Shelby Carleton and and Delia Cormier and Devon Sigurdson and John Simpson. Towards Positively Surprising Non-Player Characters in Video Games. Proceedings of the Experimental AI in Games (EXAG) Workshop at the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE). pages 7. 2017. (in press).
- Vadim Bulitko. Effects of Self-knowledge: Once Bitten Twice Shy. Proceedings of the Experimental AI in Games (EXAG) Workshop at the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE). pages 7. 2017.