Looking forward to meeting you #ECVP2025 Mainz this week, including collaborative work with @tobnie.bsky.social @dominikstrb.bsky.social @ookenfooken.bsky.social @fatatai.bsky.social @tsawallis.bsky.social @mamassian.bsky.social @guidomaiello.bsky.social @mariaeckstein.bsky.social and many others

We have an open PhD position in an exciting @dfg.de - @ageinves.bsky.social project to further develop continuous psychophysics in collaboration with Joan-Lopez Moliner.

How to infer an individual’s knowledge about the dynamics of an environment? Approximate BAMDP planning model for uncertainty over transitions & efficient replanning, as well as an approximate knowledge inference method given the behavior of an agent based on the planning model and Gibbs sampling

How to infer model parameters in sensorimotor control tasks? Dynamics may be stochastic and non-linear, the agent’s beliefs and controls may be unobserved, and beyond costs we may want to infer perceptual noises, beliefs, dynamics, and control-- this includes partial observations and unknown plant

We developed a theory of continuous-time model-based reinforcement learning generalized to arbitrary discount functions. This formulation covers non-exponential random termination times and includes solving the inverse problem of learning the discount function from decision data

Congratulations to Matthias Schultheis for defending his PhD thesis 'Inverse reinforcement learning for human decision-making under uncertainty' with distinction. Significant contributions to understanding bounded actors with inverse POMDPs for partial observabilities and non-stationary behavior

How inverse modeling can speak to algorithmic level descriptions of human behavior and the heuristics debate: What to conclude if a dynamical system model fits behavior? If it looks like online control, it is probably model-based control. Proceedings of the Annual Meeting of the Cog Sci Society

Applying inverse modeling to the continuous psychophysics paradigm: Straub, D., & Rothkopf, C. A. (2022). Putting perception into action with inverse optimal control for continuous psychophysics. eLife, 11, e76635.

Straub∗, D., Schultheis∗, M., Koeppl, H., & Rothkopf, C. A. (2023). Probabilistic inverse optimal control for non-linear partially observable systems disentangles perceptual uncertainty and behavioral costs. NeurIPS.

When dynamics and observations are non-linear, the separation principle does not hold, leading to interesting information-seeking behavior- which this method can recover! This includes partial observability from the actor's and observer's point of view, and even the dynamics can be unknown.

Inverting the classic optimal feedback control model with signal-dependent noise by Emo Todorov: Schultheis∗, M., Straub∗, D., & Rothkopf, C. A. (2021). Inverse Optimal Control Adapted to the Noise Characteristics of the Human Sensorimotor System. NeurIPS.

Performing efficient gradient-based Bayesian inference of Bayesian actor models' parameters allows principled model comparison and disentangling factors that may lead to unidentifiabilities between priors and costs

Congratulations to @dominikstrb.bsky.social for defending his PhD thesis 'Inverse normative modeling of continuous perception and action' with distinction, with significant contributions to understanding bounded actors with inverse models, reconciling normative and descriptive models of behavior

We then put our method to the test on empirical data from three sensorimotor tasks, comparing model fits with different cost functions, and it successfully explains individual behavioral patterns

Here we amortize the Bayesian actor using a neural network trained unsupervised on a broad range of parameter settings. This allows for efficient gradient-based Bayesian inference of the model’s parameters within probabilistic programming

Our latest work on understanding the behavior of bounded agents in more naturalistic tasks accepted at #ICLR2025: Inverse decision-making using neural amortized Bayesian actors, with @dominikstrb.bsky.social @tobnie.bsky.social and @jan-peters.bsky.social based on @tobnie.bsky.social MSc thesis