Modeling human-like non-rationality in social agents
Kochanowicz, Jaroslaw Slawomir
Date of Issue2018-02-09
School of Computer Science and Engineering
Rational decision making is one deliberately aimed at the realization of well-defined goals in an environment, based on a formal representation of its state and logical deduction grounded in those representations. Conversely, non-rationality is characterized by a) informal or sub-symbolic processing instead of one based on formal representations; b) associative, habitual reactions rather than step by step deduction or optimization; c) lack of explicit goals or their link to decision making. Human species is the only one capable of consistent rationality, but the scale and gravity of human non-rationality are far greater than often suspected. Cognitive biases and other non-rational influences on key decisions occur disregarding explicit reasoning or goal realization, subconsciously and often in a certainty that they did not take place. Habitual or stereotypic deviations from rationality in human cognition are well documented as non-reducible to rational pursuit of the egoistic benefit or its occasional distortion with temporary emotional excitation. Non-rationality is also not limited to the episodic emotional deviation, but is a permanent 'design feature' of human cognition, making rationalistic 'homo economicus' approach incapable of tackling the great challenge of human behavior modeling. While social psychology has systematically addressed non-rationality of human cognition for decades, it is not true for computer science, as current paradigms do not sufficiently account for the non-rational character of humans. Despite its great relevance, non-rationality is not yet properly addressed in computational models of social agents. Addressing the above problem, this thesis aims to study how various types of nonrationality can be addressed, described and modeled in the most relevant computational fields. The proposed models are implemented in the agent architectures and used in specific simulations to exemplify their ability to capture non-rationality within the personality, informal human cognition of culture and context social. To this end, thesis first presents an interdisciplinary overview of how human-like non-rationality is addressed in Cognitive Architectures (CA) , Multi-Agent Systems (MAS), Affective Computing (AF), Social Signal Processing (SSP), discussing inter-field trends and strengths and weaknesses of the specific disciplines and approaches. Based on this, elements of a paradigm of computational non-rationality modeling are proposed. Those include abandonment of the paradigmatic simplicity of MAS and closing the socio-cultural immergence-emergence loop by MAS-CAA merger. Also, an inclusion of the specific, implicit and internal culture-generating mechanisms and 'System 1' social context processing is postulated as a basis of personality and culture related non-rationality. To realize those postulates, Social Context Cognition Model (SCCM), inspired among others by the Dual Process Theory, is proposed. SCCM is a connectionist model whose activation reflects agent's subconscious social associations induced by the situation. It captures non-conceptual, habitual and associative connotations and their consequences, rather than rule and reasoning based, calculative ones. It also provides 'System 1' modulation of behavior, non-appraisal affect, action, plan or norms feasibilities, usable both in entertainment and social simulations. Furthermore, a cognitive Dramaturgical Dissonance Model (DDM) is proposed. It is inspired by a Dramaturgical Metaphor and a theory of Cognitive Dissonance and used in a simulation of 'System 1' affect generation. Next, a novel Social Context based Personality Model (SCP) is proposed based on SCCM. It is a continuation and specification of the Cognitive-Affective Personality System theory and is an alternative to the dominating ' trait models' whose drawbacks as descriptive tools are multiplied when they are used as generating mechanisms. SCP is parametrized using sociological data and used in a series of simulations including multiple agent interactions. Model's properties are compared to 'trait models' in a formal validation, proving its superiority. Finally, the issue of parametrization of socio-cognitive models, like SCCM and SCP is addressed by a semi-automatic, crowd-sourcing based parametrization system. Its presentation consists of a data acquisition system outline, including the statistical methods used and a validation, showing usability of both the semi-automatic parametrization system and the parametrized socio-cognitive models as tools for acquiring and representing human-like non-rationality.
DRNTU::Engineering::Computer science and engineering