Psychology in AI Part 2: Can Artificial Players always Beat Humans in a Complex Language Game?

Juli 14, 2023

Psychology in AI designs

        The desire in developing artificial players to compete with humans has grown recently. In language games, for example, there are certain artificial players, such as Ottho (Semeraro, de Gemmis, Lops, & Basile, 2012), that can solve a complicated language game like the Guillutine.

        The Guillutine is a language game in which players are given five words that are unrelated to each other, but strongly linked to a word that is the solution. The player then have to discover the unique solution based on those five clues (Molino et al., 2015; Semeraroet al., 2012).

        This article will discuss the strategy and mechanism of how Ottho works in solving the game from a paper entitled "An Artificial Player for a Language Game" authored by Semeraro et al. (2012), as well as discuss its limitations 

Read also:

Psychology in AI Part 1: Why Humans & Automated Cars are Not (Yet) A Good Team

1. How Ottho works?

        Ottho uses Cognitive Unit (CU) as a strategy to store the information needed in the game. CUs in Ottho operate similarly to memories in the human brain. This artificial player is inspired by how the human brain processes long-term information.

        In Ottho, CUs are information or descriptions of concepts gathered from various sources, e.g., the Encyclopedia, the Italian dictionary, proverbs, and books. They will be interconnected in the system to form a network in order to find a list of possible solutions (Semeraro et al., 2012).

        According to Semeraro et al. (2012), concepts in this artificial player are represented in two fields within the CU, i.e., the head and the body, written as CU = [head │body]. The head contains words that identify a concept within CU, whereas the body contains words that describes the concept.

        For example, Wikipedia mentions the concept of “Serendipity” as “The occurrence of unplanned fortunate discovery”, then it will be represented in CU as CU = [Serendipity │ The occurrence of unplanned fortunate discovery].


2. SAN as reasoning mechanism

        The system works by adopting Spreading Activation Network (SAN) as its reasoning mechanism. It regards the Web or knowledge sources as CU repositories that supply information during the retrieval memory process. The retrieval memory process does not begin until five hints are provided. However, for simplicity purpose, Semeraro et al. (2012) only provide two clues, i.e., Newton and sin from two sources (Dictionary and Wikipedia) for their trial.

        The system seeks the answer by establishing a network of nodes using the following approach. First, the clues will generate nodes that will be linked to CUs from CU repositories. Only those CU with the highest similarity values to the clues will be included (Semeraro et al., 2012). After all these clues have been connected with CUs, the algorithm will connect each node with keywords from the CUs' heads and bodies.

        The next process is pulse transfer. There are three steps of pulse transfer processes within the nodes: initialization, marking and firing (Semeraro et al., 2009). In initialization, all of activation values Ai(p) is set into 0, except for the clues whose activation values is set into 1.

        After that, in the marking step, each node Ai(p) ≥ F (threshold) is marked as fired. Then in the last step, the output of each node is computed as a function of its activation level. Finally, the most active nodes will be included in the list of solutions (CSL) which in the article are apple, unit, and force.


3. The limitations of SAN

        The use of an algorithm such as SAN is recognized to be beneficial in solving a complicated language game such as the Guillutine (Molino et al., 2015). The combination of multiple sources in this approach also further increases Ottho's chances of finding answers. However, if the unique answer is the intersection word of the given clues (must be related to all of the clues), the list of solutions presented in the CSL is too broad and might cause the system to become noisy.

        For example, in addition to apple, the CSL in the article offers unit and force as viable answers, since they are also considered the most active nodes (Semeraro et al., 2012). How could they potentially become candidate solutions when they are only connected to Newton from Wikipedia and Dictionary, and not part of intersections between clues (Newton and sin) like apple?


4. New reasoning mechanism

        Basile, de Germis, Lops, and Semeraro (2016) recognize the shortcoming of the mechanism they described previously in the paper in their following investigation regarding Ottho. As a result, they proposed a new mechanism approach called Reversed Path (RP), which uses CUs and SAN in their system but adds an additional condition for a word to be considered as a possible solution, where it must also be connected to all of the clues given, rather than just looking at its level of activation.

        They also add another approach to their system called Expanded Reversed Path (ERP), which works similarly to RP, but the difference is that after the SAN is completed, the system will return to the CU repositories to create new nodes in order to broaden the opportunity to find the precise solution. These novel techniques greatly improve solution precision, with ERP outperforming RP (Basile et al., 2016).

disclaimer: 

This article is a revised and improved version of my assignment for completing a course (Cognitive and Technical System) under supervision of Prof. Frank van der Velde while pursuing my master degree in Human Factors and Engineering Psychology at the University of Twente.

Reference

Basile, P., de Gemmis, M., Lops, P., & Semeraro, G. (2016). Solving a complex language game by using knowledge-based word associations discovery. IEEE Transactions on Computational Intelligence and AI in Games, 8(1), 13-26. DOI: https://doi.org/10.1109/TCIAIG.2014.2355859

Molino, P., Lops, P., Semeraro, G., de Gemmis, M., & Basile, P. (2015). Playing with knowledge: A virtual player for “Who Wants to be a Millionaire?” that leverages question answering techniques. Artificial Intelligence, 222, 157-181. DOI: https://doi.org/10.1016/j.artint.2015.02.003


Sameraro, G., de Gemmis, M., Lops, P., & Basile, P. (2012). An Artificial player for a language game. IEEE Intelligence System, 12, 36-43. Semeraro, G., Lops, P., Basile, P., de Gemmis, M. (2009). On the tip of my thought: Playing the Guillutine game. Proc. 21st Int’l Joint Conf Artificial Intelligence, 1543-1548. DOI: https://doi.org/10.1109/MIS.2011.37

You Might Also Like

0 comments

Popular Posts