Carsten Murawski's research overview
What a piece of work is a man! How noble in reason! How infinite in faculty! [...] And yet, to me, what is this quintessence of dust? Shakespeare, Hamlet
Current as of January 2026.
My research combines theoretical analysis with experiments to study how people make decisions and to find ways to improve decision-making.
To this end, I use techniques including behavioural experiments, eye-tracking, neuro-imaging, pharmacological challenges and psychophysiology in healthy and patient populations of both human and non-human animals.
The topics of my past and current research include the effects of computational complexity on decisions, (reinforcement) learning, intertemporal choice, valuation of information and preferences for early resolution of uncertainty, as well as the structure and formation of preferences.
In my translational research, I investigate consumer financial decision-making, the relation between health and decision-making as well as high-performance decision-making.
My primary research interest at the moment is in investigating how properties of biological computation, including computational constraints, in the human organism affect how we make decisions and in applying this knowledge to improve every-day decision-making.
The following provides an overview of some of my past and ongoing research in this area.
Decisions as computations
Decision-makers face two fundamental challenges: beliefs about the consequences of their decisions have to be based on limited evidence, and the evidence available needs to be perceived and processed using limited computational resources (Bossaerts, Yadav, and Murawski 2019). Over the past few years, I have been particularly interested in the second challenge.
While the brain only represents two percent of the body mass of a human organism, it consumes about 20 percent of metabolic output. This means that computation in the brain operates under a stringent resource constraint. This constraint is a driving principle of brain organisation and function. To study how computational resource constraints affect decision-making, two questions arise (Bossaerts and Murawski 2017): What computational resources are required to make a particular decision? And what are the computational capacities of decision-makers?
To investigate this problem, I make the assumption that decisions are (neuro-cognitive) computations over (neural) representations and that decision-makers are resource-limited computing systems. This allows me to formalise decisions using the theory of computation and computational complexity theory. The latter is a mathematical framework for characterising computational resource requirements.
Characterising computational resource requirements of decisions
To investigate which resources are required to make particular decisions and how different levels of resource requirements affect decision-making, we have used different approaches. In one of our early studies, with Peter Bossaerts, we asked human participants to solve different instances of the knapsack problem (Murawski and Bossaerts 2016). In the task, participants are presented with a set of items with different values and weights and have to find the combination of items that maximises total value subject to a total weight constraint. The problem participants have to solve is mathematically very similar to the consumer choice problem in economics and captures key features of many decisions people make. Critically, we picked the instances so that they varied in the number of computational resources (number of computational operations and amount of memory) required to solve them. We quantified the latter using a common algorithm for the knapsack problem. We found that the amount of time the algorithm needed to solve each instance (algorithmic complexity), that is, to make an optimal decision, predicted the amount of time participants spent on trying to solve each instance, and the probability of a participant solving an instance correctly.
A shortcoming of this study is an assumption we made in quantifying computational resource requirements: that human decision-makers use a particular, albeit common, algorithm. However, we usually don’t know which algorithms human decision-makers use (if any), and different people likely use different ones. To address this issue, we used an alternative approach called random instance analysis (typical case complexity) (Franco and Murawski 2023), which tries, using methods from computational complexity theory and statistical physics, to identify mathematical properties of instances of computational problems related to computational resource requirements that are independent of particular algorithms. In work led by Nitin Yadav, we identified a set of such properties for the knapsack problem (Yadav et al. 2020). In a follow-up study led by Pablo Franco, we showed that these properties predict both the time human participants spend on an instance and the probability that they solve an instance correctly (Franco et al. 2021). Subsequently, in a study led by Pablo Franco and Karlo Doroc, we showed that a similar set of properties predict decision time and decision quality in two other canonical types of decisions (Franco et al. 2022), based on the 3-SAT and travelling salesperson problems, suggesting that the relation between the mathematical properties of decisions on the one hand, and decision time and decision quality on the other is universal.
Recently, in a project led by Pablo Franco, we have been extending this work, using a technique called fitness landscape analysis, to characterise the computational resource requirements of individual decisions based on topological properties of the search space associated with a decision (Franco, Yadav, and Murawski 2025). We find that properties of instances associated with compute time of exact algorithms predict both decision time and decision quality of human decision-makers, at a finer level of detail than random instance analysis.
In current theoretical work, we are trying to link the different theoretical frameworks used to characterise computational resource requirements – complexity classes, random instance analysis (typical case complexity) and fitness landscape analysis. In related work, we are investigating whether properties of instances affect computational resource requirements of computing systems based on different computational models (Universal Turing Machine, quantum computer) in a similar or different way. We expect that this analysis will lead to a better understanding of the nature of computation in human organisms.
The set of studies described so far shows that mathematical properties of decisions, associated with computational resource requirements, predict both decision time and decision quality. The decisions participants made in those studies are deterministic, that is, the optimal decision (and its outcome) could always be determined with certainty, given sufficient computational resources. Most decisions, however, involve (probabilistic) uncertainty about their outcomes, requiring decision-makers to form beliefs about those outcomes given limited evidence. This means that decision-makers now face both challenges introduced above: they have to form beliefs about the consequences of their decisions based on limited evidence; and they need to process the evidence available using limited computational resources.
Limited evidence imposes an information-theoretic limit, that is, a fundamental bound on the accurary (or error) achievable by any inference procedure. It reflects the intrinsic hardness of an inference task given the amount of evidence and the underlying statistical structure. Limited computational resources impose a computational limit, that is, a fundamental bound on the accuracy achievable by a computational system, regardless of the algorithm used. It reflects the intrinsic hardness of a computational task given the computational resource limit and the structure of the computational problem. Given that inference requires computational resources, computational resource constraints lead to an information-computation gap: the information-theoretic limit is not obtainable due to a lack of sufficient computational resources.
In a recent study led by Zhongyu Xu, in collaboration with Michele Garagnani, we applied this framework to belief updating and decision-making to study how computational resource constraints affect human beliefs and decisions based on those beliefs (Zu, Murawski, and Garagnani 2025). We characterised the computational resource requirements of belief updating in a Bayesian framework. We then asked participants to update their beliefs based on new evidence and make decisions based on those beliefs. We find that the quality of their beliefs (measured as distance to the Bayesian optimum) is inversely related to the computational resources required to make the update. We also find that the quality of decisions based on those beliefs is the lower the higher the computational resource requirements of belief updating are. In other words, participants’ beliefs were often far away from the information-theoretic limit (or Bayesian optimum), and the distance to the information-theoretic limit was the larger the higher the computational resource requirements of belief formation were. We believe that these findings can explain many biases related to belief updating reported in the literature.
Taken together, the findings described above are consistent with the notion that human decision-makers are resource-limited (digital) computing systems and that decisions are based on (digital) computations. Moreover, we characterise decisions situations in which the computational resource requirements of a decision exceed the decision-maker’s computational resources. In such situations, decision-makers cannot be expected to be able to determine the preferred or optimal course of action.
A common response to this point is the claim that people, using various heuristics and other cognitive strategies, are always able to approximate the optimal course of action. This claim can be formalised using computational complexity theory. Computational problems can be categorised into different approximation classes according to, put simply, the degree to which optimal solutions of those problems can be approximated with a resource-bounded computing system. Some problems are of a nature so that no ‘approximation guarantee’ can be stated for a resource-bounded system. This means, again put simply, that for those problems, approximating the optimal solution to any degree is as computationally hard as computing the optimal solution itself. We conducted an experiment, led by Peter Bossaerts, in which we asked participants to make decisions that required them to solve computational problems from different approximation complexity classes, and found that behaviour was in line with our theoretical predictions: the harder approximation was according to theory, the worse the quality of approximation of the human participants was (Bossaerts et al. 2025).
This result, together with other findings presented above, has profound implications for decision theory as well as revealed preferences theory. The latter assumes that observed choices reflect an agent’s most preferred alternatives. However, we have shown that agents often can’t be expected to be able to identify the preferred alternative, because of insufficient computational resource requirements, and identifcation of the preferred options is more unlikely the higher the computational resource requirements of doing so are. Therefore, the assumptions on which revealed preferences theory is based become untentable, choices become difficult to interpret and any estimation of preferences becomes problematic.
More generally, our findings strongly suggest that any plausible theory of decision-making needs to take into account the computational resource requirements of decisions. Many prominent theories, from (subjective) expected utility theory to Bayesian decision theory and the heuristics and biases approach, do not.
Neural correlates of computational complexity
A significant part of my research is understanding the neural computations involved in decision-making and, in particular, how computational resource requirements affect those computations. In our first study addressing this question, led by Pablo Franco, we used ultra-high field (7 Tesla) functional magnetic resonance imaging (fMRI) to measure brain activation while participants were completing instances of the knapsack task of different levels of computational resource requirements (Franco, Bossaerts, and Murawski 2024). We identified a network of regions, including anterior insula, dorsal anterior cingulate cortex and the intra-parietal sulcus/angular gyrus, whose activation correlated with computational resource requirements. The regions have previously been associated with the encoding of uncertainty, cognitive control and mathematical reasoning, respectively. We also characterised the time courses of activation in those regions while people were making decisions as well as the connectivity between those regions. We are currently conducting follow-up studies to better understand those findings.
Computational resource allocation
A key question is how organisms allocate limited computational resources when making decisions. It is conjectured that resource allocation is based on meta-cognitive evaluations. To date, little is known about how the latter are affected by computational resource requirements. To address this question, in a study led by Xiaping Lu, we investigated the quality of self-performance estimates (SPE) of the quality of decisions with low and high computational resource requirements, in both the presence and the absence of feedback (Lu, Murawski, et al. 2025). We find that in decisions with low computational resource requirements, decision quality was high and so were estimates of self-performance. The latter were higher in the presence compared to the absence of feedback. When computational resource requirements were high, decision quality was low and so were the estimates of self-performance; however, the latter was not affected by feedback. In addition, when SPEs were high, participants paid higher prices for the opportunity to complete trials of a similar level of difficulty, establishing a link between decision quality, meta-cognition and resource allocation.
In the studies described so far, we modulated computational resource requirements of decisions while assuming that computational resources available to decision-makers are constant, that is, that they don’t vary between decisions. Assuming that computational resources in the human organism are limited, and that they are required for many different functions, it is likely that resources available to decision-making vary with context. A question of major importance to us is how the brain allocates computational resources to decisions during the decision-making process, which we are investigating in multiple ongoing studies.
One observation in our earlier studies was that many participants make decisions long before they run out of time but don’t choose the optimal course of action, a behaviour that looks like they ‘give up’ on finding the best course of action. Time courses of neural activation suggest a similar conclusion. We were interested whether we could improve decisions using so-called ‘smart drugs’, that is, drugs that act on the dopaminergic as well as (to a lesser extent) the noradrenergic systems. We hypothesised that those drugs, by increasing dopamine levels, would increase expected reward signals, which in turn would result in higher cognitive effort and better decisions. To test this hypothesis, we administered Ritalin (methylphenidate), dextroamphetamine and Modafinil in a randomised double-blinded, placebo-controlled, single-dose design involving participants without a diagnosis of a psychiatric or neurological disorder (Bowman et al. 2023). We found that effort (decision time and number of steps taken to make a decision) increased significantly, but productivity (quality of effort) decreased significantly. The latter can be attributed to increased randomness in search strategies. At the same time, productivity differences across participants decreased, even reversed, to the extent that above-average performers end up below average and vice versa.
One of the implications of our findings is that there are two dimensions to cognitive effort: extent and quality. In our study, drugs (dopamine agonists) increased the former but decreased the latter. We are conducting follow-up work to better understand the different neural pathways that affect those two dimensions of cognitive effort and their relation to decision-making.
In another study, led by Karlo Doroc, we administered a psycho-social stressor (Trier Social Stress Test) in a randomised single-blinded design before asking participant to complete instances of the knapsack task with different levels of computational resource requirements. We expected that acute stress taxes cognitive resources, thereby impairing decision-making. We found that higher cortisol levels, a marker of acute stress level, led to impaired decision quality irrespective of the level of computational resource requirements (Doroc, Yadav, and Murawski 2025). Among cortisol responders, we found that decision quality was lower while the incidence of experienced time pressure increased. Post-hoc, we found substantial deficits in decision quality when acute stress was accompanied by time pressure. Given that important life decisions, including medical and financial decisions, are often made under high levels of acute stress, our results have important implications for public policy and the design of choice architectures.
In another study, we manipulated mental (working memory) load of participants before asking them to complete the knapsack task, a cognitive resource required to complete the task (Franco et al., pre-published). Like in the previous study, we detected impaired decision quality irrespective of the level of computational resource requirements. At present, we are conducting a series of studies in which we manipulate computational resources in the brain more directly using physiological interventions.
Inter-individual differences in decision-making capacity
In addition to looking at changes in decision-making capacity within individuals, we are also interested in differences between individuals. In one recent study, led by Karlo Doroc, we examined differences in compute time and decision quality in the knapsack task between young (18-35 years old) and old participants (65 years and older) (Doroc et al., pre-published). Older participants performed significantly worse, particularly on easier trials, a deficit primarily explained by lower cognitive capacities (e.g., working memory, reasoning, set shifing) and not age, education, or motivation. Despite investing more time, older adults explored less and were more overconfident. The findings suggest that healthy cognitive ageing impairs decision-making capacity due to the decline of cognitive capacities. The findings have important implications for policy and choice architecture design, particularly in ageing populations.
In another recent study, led by Xiaping Lu, we investigated the relation between decision-making capacity and psychopathology (Lu, Keidel, et al. 2025a). Participants completed a number of trials of the knapsack task that differed in computational resource requirements as well as a set of questionnaires to assess transdiagnostic symptom dimensions related to psychopathology. We that found that overall, while performance in the knapsack task remained stable, confidence increased, leading to ‘overconfidence’ over time. Participants who scored highly on symptoms related to mood disorders (anxiety, depression) had decision performance similar to other participants but reported lower confidence. Their confidence increased over time while decision performance remained constant, resulting in miscalibration of meta-cognition (increasing level of overconfidence). On the other hand, participants who scored highly on symptoms related to compulsive behaviour and instrusive thought showed lower decision performance but higher confidence ratings, resulting in higher miscalibration of meta-cognition. In this group, confidence decreased over time, while performance remained stable, resulting in a lower degree of miscalibration (decreasing level of overconfidence) (Lu, Keidel, et al. 2025b).
The complexity of consumer decisions
In another strand of my work, I am interested in understanding people’s capacity to make important real-world decisions. One area of interest are consumer financial decisions. To this end, in a project led by Michelle Lee, we developed a framework to characterise the computational resource requirements of consumer financial decisions and applied it to credit card choice. Credit cards are among the most ubiquitous financial products, yet cause significant harm, typically due to suboptimal choices. In a combined behavioural and eye-tracking experiment, participants were asked to choose the best credit card from two alternatives. We manipulated the computational resource requirements of those decisions using features of actual credit cards offered in the market. We found that participants’ ability to identify the best card dropped to almost chance level in the more difficult cases. These findings were not driven by inattention nor due to a lack of financial literacy. Our results demonstrate how structural properties of consumer problems give rise to systematic decision errors. Importantly, the fact that even highly educated and financially literate participants struggle in this environment highlights the need to address complexity in financial decision-making
Current areas of research
Below are some of the key questions I am currently interested in:
- How does the brain allocate neuro-cognitive computational resources during decision-making?
- How are computations involved in complex decisions implemented at the neural level?
- How do the properties of biological computation differ from theoretical models of computation, in particular, Turing computation?
- What are the effects of psychiatric and neurological disorders on decision-making?
- What is the complexity of consumer decisions?
If you are interested in working with me on any of these questions as a PhD student, please get in touch.
References
Bossaerts, Peter, Juan Pablo Franco, Anthony Hsu, Carsten Murawski, and Nitin Yadav. 2025. “How Well Can Humans Approximate Optimality in Computationally Hard Problems?” Preprint.
Doroc, Karlo, Kerryn Elizabeth Pike, Juan Pablo Franco, Nitin Yadav, and Carsten Murawski. Pre-published. “Cognitive Decline, Not Age, Explains Reduced Decision-Making Capacity in Healthy Older Adults.” In preparation.
Franco, Juan Pablo, Kristian Rotaru, Elizabeth Bowman, and Carsten Murawski. Pre-published. “Pupil Diameter Reflects Computational Complexity of Decisions.” In preparation.
Franco, Juan Pablo, Nitin Yadav, and Carsten Murawski. 2025. “Using Fitness Landscape Analysis to Characterise Computational Complexity of Individual Decisions.” In preparation.
Lu, Xiaping, Kristof Keidel, Ulrich Ettinger, Carsten Murawski, and Shinsuke Suzuki. 2025a. “Psychiatric Symptoms Are Associated with Poor Performance and Enhanced Metacognition in Computationally Complex Decisions.” Preprint.
Lu, Xiaping, Kristof Keidel, Ulrich Ettinger, Carsten Murawski, and Shinsuke Suzuki. 2025b. “Statistical Inference Underlying Human Sense of Confidence in Computationally Demanding Decision-Making.” Preprint.
Yadav, Nitin, Carsten Murawski, Sebastian Sadina, and Peter Bossaerts. 2020. “Is Hardness Inherent In Computational Problems? Performance Of Human And Digital Computers On Random Instances Of The 0-1 Knapsack Problem.” In Proceedings of the 24th European Conference on Artificial Intelligence (ECAI 2020). Santiago de Compostela.
Zu, Zhongyu, Carsten Murawski, and Michele Garagnani. 2025. “The Information-Computation Gap: The Effects of Computational Complexity on Human Belief Updating.” In preparation.