Research interests

Living beings have the marvelous capacity to make decisions and plan their actions in a sophisticated way. This capacity is a direct consequence of their evolution and the environments where they evolved. This historical trajectory provides constraints not only to the types of decisions that organisms can make, but above all it continuously influences their interests in a way that we cannot fully understand their decisions without taking into acount their evolutionary ecologies. I am interested in studying decisions and actions as a consequence of goals generated autonomously and continuously by living beings, under the light of their eco-evo-developmental trajectories.

The main questions that haunt me day and night:

• It seems apparent that agents do things for reasons. All the algorithms and mechanisms we hypothesize for decision making assume certain reasons for behavior. In my view, it is of utmost importance we aim to understand those reasons, or else our algorithmic and mechanistic understanding will be skewed. In short, what are agents trying to achieve? Why and how does an agent build those objectives? And how do those objectives change with context?
• If agency is an evolved trait, there should be a continuous trajectory of action abstraction. In fields like RL, the idea of action is almost a starting point, yet for me it is an end goal: how do actions emerge in complex systems that sustain themselves? How are action spaces formed and maintained to produce behavior? In other (more philosophical) words, where is the locus of agency and how plastic is it?

Two projects I have worked on in the past years: the Maximum Occupancy Principle and the Breadth-Depth dilemma.

Behavior as occupancy of action-state path space

In theories of behavior, it is usually assumed that agents maximize reward. From an algorithmic point of view, this makes sense: reward is any measure of goals that agents are interested in achieving. Its source can be external (e.g. training a robot to do something), or internal (e.g. homeostasis) to the agent. However, inferring reward signals from natural behavior and designing reward signals for artifical agents can be problematic.

In this work, we propose a simple principle of behavior: occupying action-state path space. “Reward” in the usual sense becomes a means to achieve this goal, conceptualizing it as transitions between internal states. From this principle, we obtain diverse, complex behaviors that simple reward maximizing agents with endowed variability do not display. See publications for more information.

Breadth-depth dilemma

The breadth-depth dilemma is a bounded-rationality model to investigate the influence of sampling resource constraints (e.g. time, precision, compute) on decision-making strategies. We have found different regimes of optimality that depend on the amount of sampling resources, both for discrete and continuous resources. In both of these cases, the optimal number of options to sample is a (or very close to) a power law for large amount of resources. Funny things happen at the transition!