Humans balance breadth and depth: Near-optimal performance in many-alternative decision making

Many everyday life decisions require allocating finite resources, such as attention or time, to examine multiple available options, like choosing an online food supplier. In these cases, our search resources can be spread across many options (breadth) or focused on a few of them (depth). Whilst theoretical work has described how finite resources should be allocated to maximise utility in these problems, evidence about how humans balance breadth and depth is lacking. We introduce a novel experimental paradigm where humans make a many-alternative decision under finite resources. In an imaginary scenario, participants allocate a finite budget to sample amongst multiple apricot suppliers in order to estimate the quality of their fruits, and ultimately choose the best one. We found that at low budget capacity participants sample as many suppliers as possible, and thus prefer breadth, whereas at high capacities participants sample just a few chosen alternatives in depth, and intentionally ignore the rest. The number of alternatives sampled increases with capacity following a power law with an exponent close to 0.75. In richer environments, where good outcomes are more likely, humans further favour depth. Participants deviate from optimality and tend to allocate capacity amongst the selected alternatives more homogeneously than it would be optimal, but the impact on the outcome is small. Overall, our results undercover a rich phenomenology of close-to-optimal behaviour and biases in complex choices.

The Applicability of Self-Play Algorithms to Trading and Forecasting Financial Markets

The central research question to answer in this study is whether the AI methodology of Self-Play can be applied to financial markets. In typical use-cases of Self-Play, two AI agents play against each other in a particular game, e.g., chess or Go. By repeatedly playing the game, they learn its rules as well as possible winning strategies. When considering financial markets, however, we usually have one player—the trader—that does not face one individual adversary but competes against a vast universe of other market participants. Furthermore, the optimal behaviour in financial markets is not described via a winning strategy, but via the objective of maximising profits while managing risks appropriately. Lastly, data issues cause additional challenges, since, in finance, they are quite often incomplete, noisy and difficult to obtain. We will show that academic research using Self-Play has mostly not focused on finance, and if it has, it was usually restricted to stock markets, not considering the large FX, commodities and bond markets. Despite those challenges, we see enormous potential of applying self-play concepts and algorithms to financial markets and economic forecasts.

Predictive eye movements are adjusted in a Bayes-optimal fashion in response to unexpectedly changing environmental probabilities

This paper examines the application of active inference to naturalistic visuomotor control. Active inference proposes that actions serve to minimise future prediction errors and are dynamically adjusted according to uncertainty about sensory information, predictions, or the environment. We investigated whether predictive gaze behaviours are indeed adjusted in this Bayes-optimal fashion during a virtual racquetball task. In this task, participants intercepted bouncing balls with varying levels of elasticity, under conditions of high and low environmental volatility. Participants’ gaze patterns differed between stable and volatile conditions in a manner consistent with generative models of Bayes-optimal behaviour. Partially observable Markov models also revealed an increased rate of associative learning in response to unpredictable shifts in environmental probabilities, although there was no overall effect of volatility on this parameter. Findings extend active inference frameworks into complex and unconstrained visuomotor tasks and present important implications for a neurocomputational understanding of the visual guidance of action.

Experimental Verification of Self-Adapting Data-Driven Controllers in Active Distribution Grids

Lately, data-driven algorithms have been proposed to design local controls for Distributed Generators (DGs) that can emulate the optimal behaviour without any need for communication or centralised control. The design is based on historical data, advanced off-line optimization techniques and machine learning methods, and has shown great potential when the operating conditions are similar to the training data. However, safety issues arise when the real-time conditions start to drift away from the training set, leading to the need for online self-adapting algorithms and experimental verification of data-driven controllers. In this paper, we propose an online self-adapting algorithm that adjusts the DG controls to tackle local power quality issues. Furthermore, we provide experimental verification of the data-driven controllers through power Hardware-in-the-Loop experiments using an industrial inverter. The results presented for a low-voltage distribution network show that data-driven schemes can emulate the optimal behaviour and the online modification scheme can mitigate local power quality issues.

Behavioural flexibility and reputation formation

Limited flexibility in behaviour gives rise to behavioural consistency, so that past behaviour is partially predictive of current behaviour. The consequences of limits to flexibility are investigated in a population in which pairs of individuals play a game of trust. The game can either be observed by others or not. Reputation is based on trustworthiness when observed and acts as a signal of behaviour in future interactions with others. Individuals use the reputation of partner in deciding whether to trust them, both when observed by others and when not observed. We explore the effects of costs of exhibiting a difference in behaviour between when observed and when not observed (i.e. a cost of flexibility). When costs are low, individuals do not attempt to signal that they will later be trustworthy: their signal should not be believed since it will always pay them to be untrustworthy if trusted. When costs are high, their local optimal behaviour automatically acts as an honest signal. At intermediate costs, individuals are very trustworthy when observed in order to convince others of their trustworthiness when unobserved. It is hypothesized that this type of strong signalling might occur in other settings.

Economic Analysis of Accident Law: A New Liability Rule that Induces Socially Optimal Behaviour in Case of Limited Information

In accident law, we seek a liability rule that will induce both the parties to adopt socially optimal levels of precaution. Economic analysis, however, shows that none of the commonly used liability rules induce both parties to adopt optimal levels, if courts have access only to ‘Limited Information’ on. In such a case, it has also been established (K. (2006). Efficiency of liability rules: a reconsideration. J. Int. Trade Econ. Dev. 15: 359–373) that no liability rule based on cost justified untaken precaution as a standard of care can be efficient. In this paper, we describe a two-step liability rule: the rule of negligence with the defence of relative negligence. We prove that this rule has a unique Nash equilibrium at socially optimal levels of care for the non-cooperative game, and therefore induces both parties to adopt socially optimal behaviour even in case of limited information.

Alarm Systems in Buildings – Investigation of Effectiveness Depending on Building Type and User Profile

Alarm systems are used to warn passengers and employees in railway stations, airports, logistics facilities and administrative buildings of hazards; in this way, they are instructed to rescue themselves. A large online survey conducted in Germany in 2018 has shown that often sounding the alarm does not have the desired effect. The research presented in this article investigates dependence of an alarm effectiveness on type of building, user profile and type of the alarm system. In this paper, general and building-specific measures, using which the effective sounding of the alarm and optimal behaviour of the building users in the case of danger can be achieved, are presented.

Scalable Methods for Computing State Similarity in Deterministic Markov Decision Processes

We present new algorithms for computing and approximating bisimulation metrics in Markov Decision Processes (MDPs). Bisimulation metrics are an elegant formalism that capture behavioral equivalence between states and provide strong theoretical guarantees on differences in optimal behaviour. Unfortunately, their computation is expensive and requires a tabular representation of the states, which has thus far rendered them impractical for large problems. In this paper we present a new version of the metric that is tied to a behavior policy in an MDP, along with an analysis of its theoretical properties. We then present two new algorithms for approximating bisimulation metrics in large, deterministic MDPs. The first does so via sampling and is guaranteed to converge to the true metric. The second is a differentiable loss which allows us to learn an approximation even for continuous state MDPs, which prior to this work had not been possible.