Studying sturdy controllers that work throughout many partially observable environments

In clever techniques, functions vary from autonomous robotics to predictive upkeep issues. To regulate these techniques, the important facets are captured with a mannequin. After we design controllers for these fashions, we nearly at all times face the identical problem: uncertainty. We’re not often capable of see the entire image. Sensors are noisy, fashions of the system are imperfect; the world by no means behaves precisely as anticipated.

Think about a robotic navigating round an impediment to succeed in a “aim” location. We summary this situation right into a grid-like setting. A rock could block the trail, however the robotic doesn’t know precisely the place the rock is. If it did, the issue could be fairly straightforward: plan a route round it. However with uncertainty in regards to the impediment’s place, the robotic should study to function safely and effectively regardless of the place the rock seems to be.

This straightforward story captures a much wider problem: designing controllers that may deal with each partial observability and mannequin uncertainty. On this weblog publish, I’ll information you thru our IJCAI 2025 paper, “Strong Finite-Reminiscence Coverage Gradients for Hidden-Mannequin POMDPs”, the place we discover designing controllers that carry out reliably even when the setting might not be exactly identified.

When you may’t see all the pieces

When an agent doesn’t absolutely observe the state, we describe its sequential decision-making drawback utilizing a partially observable Markov determination course of (POMDP). POMDPs mannequin conditions wherein an agent should act, primarily based on its coverage, with out full data of the underlying state of the system. As a substitute, it receives observations that present restricted details about the underlying state. To deal with that ambiguity and make higher choices, the agent wants some type of reminiscence in its coverage to recollect what it has seen earlier than. We usually signify such reminiscence utilizing finite-state controllers (FSCs). In distinction to neural networks, these are sensible and environment friendly coverage representations that encode inner reminiscence states that the agent updates because it acts and observes.

From partial observability to hidden fashions

Many conditions not often match a single mannequin of the system. POMDPs seize uncertainty in observations and within the outcomes of actions, however not within the mannequin itself. Regardless of their generality, POMDPs can’t seize units of partially observable environments. In actuality, there could also be many believable variations, as there are at all times unknowns — completely different impediment positions, barely completely different dynamics, or various sensor noise. A controller for a POMDP doesn’t generalize to perturbations of the mannequin. In our instance, the rock’s location is unknown, however we nonetheless need a controller that works throughout all attainable areas. This can be a extra practical, but in addition a tougher situation.

To seize this mannequin uncertainty, we launched the hidden-model POMDP (HM-POMDP). Reasonably than describing a single setting, an HM-POMDP represents a set of attainable POMDPs that share the identical construction however differ of their dynamics or rewards. An necessary reality is {that a} controller for one mannequin can be relevant to the opposite fashions within the set.

The true setting wherein the agent will finally function is “hidden” on this set. This implies the agent should study a controller that performs properly throughout all attainable environments. The problem is that the agent doesn’t simply need to cause about what it could’t see but in addition about which setting it’s working in.

A controller for an HM-POMDP should be sturdy: it ought to carry out properly throughout all attainable environments. We measure the robustness of a controller by its sturdy efficiency: the worst-case efficiency over all fashions, offering a assured decrease sure on the agent’s efficiency within the true mannequin. If a controller performs properly even within the worst case, we might be assured it’ll carry out acceptably on any mannequin of the set when deployed.

In direction of studying sturdy controllers

So, how can we design such controllers?

We developed the sturdy finite-memory coverage gradient rfPG algorithm, an iterative method that alternates between the next two key steps:

• Strong coverage analysis: Discover the worst case. Decide the setting within the set the place the present controller performs the worst.
• Coverage optimization: Enhance the controller for the worst case. Regulate the controller’s parameters with gradients from the present worst-case setting to enhance sturdy efficiency.

Over time, the controller learns sturdy habits: what to recollect and how you can act throughout the encountered environments. The iterative nature of this method is rooted within the mathematical framework of “subgradients”. We apply these gradient-based updates, additionally utilized in reinforcement studying, to enhance the controller’s sturdy efficiency. Whereas the small print are technical, the instinct is straightforward: iteratively optimizing the controller for the worst-case fashions improves its sturdy efficiency throughout all of the environments.

Underneath the hood, rfPG makes use of formal verification methods carried out within the device PAYNT, exploiting structural similarities to signify giant units of fashions and consider controllers throughout them. Thanks to those developments, our method scales to HM-POMDPs with many environments. In follow, this implies we are able to cause over greater than 100 thousand fashions.

What’s the impression?

We examined rfPG on HM-POMDPs that simulated environments with uncertainty. For instance, navigation issues the place obstacles or sensor errors various between fashions. In these exams, rfPG produced insurance policies that weren’t solely extra sturdy to those variations but in addition generalized higher to utterly unseen environments than a number of POMDP baselines. In follow, that means we are able to render controllers sturdy to minor variations of the mannequin. Recall our operating instance, with a robotic that navigates a grid-world the place the rock’s location is unknown. Excitingly, rfPG solves it near-optimally with solely two reminiscence nodes! You possibly can see the controller beneath.

By integrating model-based reasoning with learning-based strategies, we develop algorithms for techniques that account for uncertainty moderately than ignore it. Whereas the outcomes are promising, they arrive from simulated domains with discrete areas; real-world deployment would require dealing with the continual nature of varied issues. Nonetheless, it’s virtually related for high-level decision-making and reliable by design. Sooner or later, we are going to scale up — for instance, by utilizing neural networks — and goal to deal with broader lessons of variations within the mannequin, comparable to distributions over the unknowns.

Need to know extra?

Thanks for studying! I hope you discovered it attention-grabbing and bought a way of our work. You could find out extra about my work on marisgg.github.io and about our analysis group at ai-fm.org.

This weblog publish relies on the next IJCAI 2025 paper:

• Maris F. L. Galesloot, Roman Andriushchenko, Milan Češka, Sebastian Junges, and Nils Jansen: “Strong Finite-Reminiscence Coverage Gradients for Hidden-Mannequin POMDPs”. In IJCAI 2025, pages 8518–8526.

For extra on the methods we used from the device PAYNT and, extra usually, about utilizing these methods to compute FSCs, see the paper beneath:

• Roman Andriushchenko, Milan Češka, Filip Macák, Sebastian Junges, Joost-Pieter Katoen: “An Oracle-Guided Strategy to Constrained Coverage Synthesis Underneath Uncertainty”. In JAIR, 2025.

If you happen to’d wish to study extra about one other means of dealing with mannequin uncertainty, take a look at our different papers as properly. As an example, in our ECAI 2025 paper, we design sturdy controllers utilizing recurrent neural networks (RNNs):

• Maris F. L. Galesloot, Marnix Suilen, Thiago D. Simão, Steven Carr, Matthijs T. J. Spaan, Ufuk Topcu, and Nils Jansen: “Pessimistic Iterative Planning with RNNs for Strong POMDPs”. In ECAI, 2025.

And in our NeurIPS 2025 paper, we research the analysis of insurance policies:

• Merlijn Krale, Eline M. Bovy, Maris F. L. Galesloot, Thiago D. Simão, and Nils Jansen: “On Evaluating Insurance policies for Strong POMDPs”. In NeurIPS, 2025.