Discovering Rational Heuristics for Risky Choice

For computationally limited agents such as humans, perfectly rational decision-making is almost always out of reach. Instead, people may rely on computationally frugal heuristics that usually yield good outcomes. Although previous research has identified many such heuristics, discovering good heuristics and predicting when they will be used remains challenging. Here, we present a machine learning method that identifies the best heuristics to use in any given situation. To demonstrate the generalizability and accuracy of our method, we compare the strategies it discovers against those used by people across a wide range of multi-alternative risky choice environments in a behavioral experiment that is an order of magnitude larger than any previous experiments of its type. Our method rediscovered known heuristics, identifying them as rational strategies for specific environments, and discovered novel heuristics that had been previously overlooked. Our results show that people adapt their decision strategies to the structure of the environment and generally make good use of their limited cognitive resources, although they tend to collect too little information and their strategy choices do not always fully exploit the structure of the environment.

Model-based prioritization for acquiring protection

Protection, or the mitigation of harm, often involves the capacity to prospectively plan the actions needed to combat a threat. The computational architecture of decisions involving protection remains unclear, as well as whether these decisions differ from other positive prospective actions. Here we examine effects of valence and context by comparing protection to reward, which occurs in a different context but is also positively valenced, and punishment, which also occurs in an aversive context but differs in valence. We applied computational modeling across three independent studies (Total N=600) using five iterations of a ‘two-step’ behavioral task to examine model-based reinforcement learning for protection, reward, and punishment in humans. Decisions motivated by acquiring safety via protection evoked a higher degree of model-based control than acquiring reward and avoiding punishment, with no significant differences in learning rate. The context-valence asymmetry characteristic of protection increased deployment of flexible decision strategies, suggesting model-based control depends on the context in which outcomes are encountered as well as the valence of the outcome.

Simulated visual hallucinations in virtual reality enhance cognitive flexibility

Historically, psychedelic drugs are known to modulate cognitive flexibility, a central aspect of cognition permitting adaptation to changing environmental demands. Despite proof suggesting phenomenological similarities between artificially-induced and actual psychedelic altered perception, experimental evidence is still lacking about whether the former is also able to modulate cognitive flexibility. To address this, we measure participants' cognitive flexibility through behavioral tasks after the exposure to virtual reality panoramic videos and their hallucinatory-like counterparts generated by the DeepDream algorithm. Results show that the estimated semantic network has a flexible structure when preceded by altered videos. Crucially, following the simulated psychedelic exposure, individuals also show an attenuated contribution of the automatic process and chaotic dynamics underlying the decision process. This suggests that simulated altered perceptual phenomenology enhances cognitive flexibility, presumably due to a reorganization in the cognitive dynamics that facilitates the exploration of uncommon decision strategies and inhibits automated choices.

Prefrontal top-down projections control context-dependent strategy selection

The rules governing behavior often vary with behavioral contexts. As a consequence, an action rewarded in one context may be discouraged in another. Animals and humans are capable of switching between behavioral strategies under different contexts and acting adaptively according to the variable rules, a flexibility that is thought to be mediated by the prefrontal cortex (PFC)1-4. However, how the PFC orchestrates context-dependent switch of strategies remains unclear. Here we show that pathway-specific projection neurons in the medial PFC (mPFC) differentially contribute to context-instructed strategy selection. In a decision-making task in which mice have been trained to flexibly switch between a previously established rule and a newly learned rule in a context-dependent manner, the activity of mPFC neurons projecting to the dorsomedial striatum encodes the contexts, and further represents decision strategies conforming to the old and new rules. Moreover, the activity of these neuron is required for context-instructed strategy selection. In contrast, the activity of mPFC neurons projecting to the ventral midline thalamus does not discriminate between the contexts, and represents the old rule even if mice have adopted the new one; furthermore, these neurons act to prevent the strategy switch under the new rule. Our results suggest that the mPFC→striatum pathway promotes flexible strategy selection guided by contexts, whereas the mPFC→thalamus pathway favors fixed strategy selection by preserving old rules. Balanced activity between the two pathways may be critical for adaptive behaviors.

Driver-like decision-making method for vehicle longitudinal autonomous driving based on deep reinforcement learning

Decision-making is one of the key parts of the research on vehicle longitudinal autonomous driving. Considering the behavior of human drivers when designing autonomous driving decision-making strategies is a current research hotspot. In longitudinal autonomous driving decision-making strategies, traditional rule-based decision-making strategies are difficult to apply to complex scenarios. Current decision-making methods that use reinforcement learning and deep reinforcement learning construct reward functions designed with safety, comfort, and economy. Compared with human drivers, the obtained decision strategies still have big gaps. Focusing on the above problems, this paper uses the driver’s behavior data to design the reward function of the deep reinforcement learning algorithm through BP neural network fitting, and uses the deep reinforcement learning DQN algorithm and the DDPG algorithm to establish two driver-like longitudinal autonomous driving decision-making models. The simulation experiment compares the decision-making effect of the two models with the driver curve. The results shows that the two algorithms can realize driver-like decision-making, and the consistency of the DDPG algorithm and human driver behavior is higher than that of the DQN algorithm, the effect of the DDPG algorithm is better than the DQN algorithm.

Boosting promotes advantageous risk-taking

Due to the prevalence and importance of choices with uncertain outcomes, it is essential to establish what interventions improve risky decision-making, how they work, and for whom. Two types of low-intensity behavioural interventions are promising candidates: nudges and boosts. Nudges guide people to better decisions by altering how a choice is presented, without restricting any options or modifying the underlying payoff matrix. Boosts, on the other hand, teach people decision strategies that focus their attention on key aspects of the choice, which allows them to make more informed decisions. A recent study compared these two types of interventions and found that boosts worked better for risky choices aimed at maximising gains, whereas nudges worked best for choices aimed at minimising losses. Though intriguing, these findings could not be easily interpreted because of a limitation in the items used. Here we replicate that study, with an extended item set. We find that boosts work by promoting risk-taking when it is beneficial, whereas nudges have a consistent (lesser) impact, regardless of whether risk-taking is beneficial or not. These results suggest that researchers and policymakers should consider the base rate risk propensity of the target population when designing decision-support systems.

Experts use base rates in real-world sequential decisions

Human behavior is often assumed to be irrational, full of errors, and affected by cognitive biases. One of these biases is base-rate neglect, which happens when the base rates of a specific category are not considered when making decisions. We argue here that while naïve subjects demonstrate base-rate neglect in laboratory conditions, experts tested in the real world do use base rates. Our explanation is that lab studies use single questions, whereas, in the real world, most decisions are sequential in nature, leading to a more realistic test of base-rate use. One decision that lends itself to testing base-rate use in real life occurs in beach volleyball—specifically, deciding to whom to serve to win the game. Analyzing the sequential choices in expert athletes in more than 1,300 games revealed that they were sensitive to base rates and adapted their decision strategies to the performance of the opponent. Our data describes a threshold at which players change their strategy and use base rates. We conclude that the debate over whether decision makers use base rates should be shifted to real-world tests, and the focus should be on when and how base rates are used.

Decision Strategies while Intoxicated relate to Alcohol-Impaired Driving Attitudes and Intentions

Objective: Approximately 28 million individuals engage in alcohol-impaired driving (AID) every year. This study investigated individuals’ AID decision making strategies under intoxication, their variability across the breath alcohol concentration curve (BrAC), and the association between strategy and AID attitudes and intentions. Method: 79 adults (23.9 years, 57% women) who drank alcohol ≥2 days per week and lived >2 miles away from their typical drinking locations completed an alcohol administration protocol and AID decision making task. AID attitudes, intentions, and behaviors were assessed repeatedly across the BrAC curve. Bayesian cognitive modeling identified decision strategies used by individuals on the AID decision making task, revealing whether alcohol consumption level and/or ride service cost factored into individuals’ decisions to drive while impaired or obtain a ride. Additional analyses tested whether AID attitudes and intentions were related to individuals’ decision strategies. Results: Two decision strategies were examined on the ascending and descending limb of the BrAC curve: compensatory (both consumption level and ride service cost factored into AID decisions) and non-compensatory (only consumption level factored into AID decisions). Switching to a compensatory strategy on the descending limb was associated with lower perceived intoxication, perceiving AID as less dangerous, and being willing to drive above the legal BrAC limit. Conclusions: Results suggest that risk for engaging in AID is higher for those using a cost-sensitive, compensatory strategy when making AID decisions under intoxication. Future research is needed to test whether AID countermeasures (e.g., subsidized ride services) are differentially effective according to decision strategy type.