Statement of purpose: cultural knowledge evolution in 10+3 questions

What does mOeX do?

mOeX aims to understand the principles guiding knowledge evolution in social agents. It considers agents behaving according to their knowledge and beliefs.

In its whole generality, this applies to human beings, possibly animals, as well as software agents, possibly neuronal. But, we study them chiefly in a well-controlled computer science context using formal knowledge representations.

Why does mOeX do this?

There are two benefits in doing this:

• improving the understanding of knowledge evolution principles which are, by far, unknown,
• implementing computer systems whose capability to evolve their knowledge is known.

What is this useful for?

This is useful, because our societies are increasingly producing knowledge-based systems which have to interact with their environments including other agents and human beings. If such systems are not able to evolve their knowledge through these interactions, they will be inadequate and may cause accidents.

This applies to social robots designed to help elderly people or smart cities monitoring the social behaviour to facilitate it.

Although we do not develop applications ourselves, we are open to collaborations.

Which theories does mOeX take inspiration from?

We adopt an evolutionary perspective to this problem. Our work is based on cultural evolution which applies evolution theory to culture. It has proven useful for explaining long term social phenomena and is wide spread in social sciences and the humanities, from anthropology to social psychology.

We apply it to knowledge and beliefs considered as culture, in a computational context. More specifically, we aim at assessing the global properties of knowledge evolved through local adaptation. (This local/global opposition is understood with respect to society —individual/population—, space —close/far— and time —short-term/long-term.)

We assume that cultural evolution can be seen as a computational phenomenon, in the sense that cultural evolution applies to computational entities (this is artificial cultural evolution) and that a computational approach can help understanding cultural evolution (this is agent-based simulation of cultural evolution).

Which methodological approaches does mOeX use?

We perform:

• Experimental work using multi-agent simulation: we develop agents and adaptation operators, simulate their behaviour and observe the properties of the obtained knowledge;
• Theoretical work using logics: we model adaptation mechanisms logically and prove their properties.

As a third line of work, we would like to improve connections with social sciences and the humanities.

Which techniques does mOeX build on?

We use:

• Symbolic knowledge representation, in particular the ability to express knowledge in well-defined logic identifying consequences and consistency;
• Dynamic operations manipulating knowledge, such as belief revision, inductive learning or dynamic modalities;
• Multi-agent simulation, in which agents playing interaction-games involving their knowledge are monitored.

More precisely, knowledge is subject to internal constraints, imposed by logical coherence, and external constraints, imposed by interaction with the environment and others.

We study populations of agents under these constraints. Interactions are carried out through precisely specified modalities (which may involve direct knowledge exchange, talking, acting together or in presence). After interacting, when they discover that their knowledge is not adequate, agents adapt their knowledge to the situation.

The repetition of such adaptations constitute the evolution of the knowledge of a population of agents to their environment.

What has been achieved so far?

We have performed:

• Simulation of agent populations improving their knowledge (ontology, alignments) under various circumstances. In particular, we highlighted operators allowing agents to improve their knowledge and preserve its diversity through aiming at agreeing.
• Study of the transmission modes across generations of agents. We showed that vertical transmission provided enough variation to further improve knowledge without constraining reproduction.
• Studies of the properties of adaptation operators in dynamic epistemic logic. We proved epistemic properties of the operators used in actual experiments with respect to a dynamic epistemic logic models (correctness, incompleteness, redundancy) and compared what could be obtained by both approaches (agents are robust to inconsistency).

Like any good research work, as it answers some questions, it brings new ones, eventually more fundamental.

Which questions did this work brought?

Interesting questions that we would like to consider in depth now are:

• the relations between cultural knowledge evolution and reinforcement learning;
• the relations between cultural knowledge evolution and belief revision.

What is still on your plate?

There are two long-term threads of questions which are related to the dynamics of the context in which agents evolve. This occurs when agent populations:

1. experience changes in the environment in which they operate, or
2. encounter other populations with which they have to communicate.

Adding population encounter and environment change.

Is that all?

These are the main topics. But we also consider pluripotent agents, diversity and resilience, creativity and self-motivation, value-based evolution operators, knowledge propagation and network structure as well as links with replicator-interactor.

Does it raise ethical problems?

Not in itself.

But as any kind of knowledge it may be used for good or bad. It may be used to understand how populations may better align their knowledge to less conflicting relations. It may also be used to manipulate them in order to reach a specific kind of knowledge.

More down to earth, it may be used to design robots to help elderly people in an acceptable way. Doing so, this may encourage our society to leave elderly people in the care of robots...

Isn't it good old-fashioned AI ignoring recent advances in deep learning?

We develop our work with symbolic techniques which have the advantage of being more directly understandable than neuro-statistic approaches. However, the raised problems occur for such systems and may be handled along similar principles. Hence, exploring the interwinning of symbolic constraints and subsymbolic learning is a natural extension.

Are things that simple?

They ain't.