Towards agents with human-like decisions under uncertainty
2013; Wiley; Volume: 35; Issue: 35 Linguagem: Inglês
ISSN
1551-6709
AutoresNuno C. Marques, Francisco S. Melo, Samuel Mascarenhas, João Dias, Rui Prada, Ana Paiva,
Tópico(s)Cognitive Science and Mapping
ResumoTowards agents with human-like decisions under uncertainty 1 Nuno Marques (nunocm@gmail.com) Francisco Melo (fmelo@inesc-id.pt) Samuel Mascarenhas (samuel.fm@gmail.com) Jo˜ao Dias (joao.dias@gaips.inesc-id.pt) Rui Prada (rui.prada@gaips.inesc-id.pt) Ana Paiva (ana.paiva@inesc-id.pt) INESC-ID, Instituto Superior T´ecnico Av. Prof. Cavaco Silva, TagusPark 2780-990 Porto Salvo, Portugal Abstract Creating autonomous virtual agents capable of exhibiting human-like behaviour under uncertainty is becoming increas- ingly relevant, for instance in multi-agent based simulations (MABS), used to validate social theories, and also as intelli- gent characters in virtual training environments (VTEs). The agents in these systems should not act optimally; instead, they should display intrinsic human limitations and make judge- ment errors. We propose a Belief-Desire-Intention (BDI) based model which allows for the emergence of uncertainty re- lated biases during the agent’s deliberation process. To achieve it, a probability of success is calculated from the agent’s beliefs and attributed to each available intention. These probabilities are then combined with the intention’s utility using Prospect Theory, a widely validated descriptive model of human deci- sion. We also distinguish risk from ambiguity, and allow for individual variability in attitudes towards these two types of uncertainty through the specification of indices. In a travelling scenario, we demonstrate how distinct, more realistic agent be- haviours can be obtained by applying the proposed model. Keywords: Intelligent agents; Decision making; Cognitive bi- ases Introduction Uncertainty is a natural part of our world. No one can claim to know everything, no one can predict the future. We deal with uncertainty on our everyday lives and our behaviour is constantly influenced by it, even if we do not always realize it. However, in the context of virtual agents, uncertainty has usually been seen as a problem that the agent must overcome (eg. planning Peot & Smith, 1992), and thus most existing systems are aimed at achieving optimal agent behaviour un- der these conditions. Our approach is different, in which we acknowledge the of- ten sub-optimal, even “irrational” behaviour of humans when confronted with uncertain situations. These decision biases and judgement errors have been extensively studied and are supported by a wealth of empirical evidence (eg. Kahneman & Tversky, 1979; Camerer & Ho, 1994). We propose an agent model based on the classical Belief-Desire-Intention (BDI) paradigm, which seeks to integrate in the agent’s de- liberation process these deviations from rational behaviour. 1 This work was partially supported by the Portuguese Fundac¸ a ˜ o para a Ciˆencia e Tecnologia under project PEst- OE/EEI/LA0021/2011. (INESC-ID multiannual funding) through the PIDDAC Program funds. Additionally, it was funded by the National Project SEMIRA (ERA-Compl/0002/2009), and European Project eCute (ICT-5-4.2 257666) projects. Agents with the aforementioned characteristics can be spe- cially useful for Multi-Agent Based Simulations (MABS) (Davidsson, 2001). In these systems, human behaviour is modelled at the individual (agent) level, and the resulting structure is analysed after it emerges from the agent inter- actions. Typically, MABS have been used to validate so- cial theories (eg. Davidsson, 2002). The inclusion of un- certainty is of special importance in market simulations, as it strongly impacts the decisions of the agent (Arthur, 1991). From socio-cultural research, the Uncertainty Avoidance di- mension of human cultures, identified by Hofstede (Hofstede, 2001), is another example where these agents could be used in the context of MABS. Our solution is also relevant for use in serious games, particularly virtual training environments. As these simulation often focus on social and communication aspects (eg. Johnson & Valente, 2009; Kim et al., 2009), it is increasingly important to embed the virtual characters with human-like behaviour. This paper is organized as follows. We start by giving a possible definition of uncertainty and describing Prospect Theory, and follow with work related to ours. Then we present the model, and demonstrate it using an example sce- nario. Finally we discuss future improvements. Background In tackling the effects of uncertainty, one should first have an accurate definition of the term. However, this is not an easy task because different research fields or problem approaches use it with different meanings. One important step is distinguishing uncertainty from the closely related concept of risk. In a decision context, the later refers to choices involving known chances (eg. a spin of a roulette wheel). However, uncertainty arises in a decisions involving personal opinions (eg. betting on what football team will win a game). Moreover, uncertainty has distinct facets (Smithson, 2008): epistemic randomness or risk un- certainty is the subjective counterpart of risk, and is usually represented by subjective probabilities; ambiguity, which re- sults from overlapping beliefs (i.e, strong reasons to believe and not believe) or uncertainty about probabilities (second order uncertainty); and vagueness, reflected by fuzzy state- ments (eg. “John is tall” — what does “tall” mean?). The topic of how humans choose (or should choose) under
Referência(s)