Human Intuition Proves Stronger than Algorithms: Game Theory Tournament Held at HSE University in Perm

© iStock

Researchers from the International Laboratory of Intangible-driven Economy (Perm) and the HSE Laboratory of Sports Studies, together with mathematician and science populariser Alexey Savvateev, organised a game theory tournament entitled ‘The Election Race.’ Participants competed both against one another and against artificial intelligence. For now, humans have managed to gain the upper hand and propose more effective strategies.

Scientists are actively studying how artificial intelligence algorithms may influence decision-making in various fields. One possible approach is to conduct experiments involving decision-making tasks in which both humans and large language models take part. Their interaction makes it possible to understand who grasps the task context more deeply and how different types of agents respond to one another.

Economists at HSE University organised a game theory tournament involving both human participants and large language models. As the basis for the task, the organisers used the Colonel Blotto game (for more details, see the game rules in Russian), but placed it in a different context so that it would be less recognisable and, particularly for AI, there would be no opportunity to find optimal solutions online.

Candidates for president of a fictional country were required to make 100 campaign visits across nine states. In each state, the candidate who visited more frequently was declared the winner. In other words, the only factor determining voters’ choice was the frequency of meetings with a candidate. The presidency was awarded to the candidate who won in the greater number of states. The players’ task was to allocate their visits strategically to claim victory.

Dmitry Dagaev

‘The Colonel Blotto game—or in our version ‘The Election Race’ —is interesting because there are no dominant strategies or Nash equilibrium in pure strategies. This means that when choosing a strategy, a player cannot rely on any universally optimal solution. In order to win, it is crucial to understand who your opponent is and how to outperform them. A task with many opponents (and in our tournament there were more than 200 human participants and the same number of AI strategies) is even more complex: one has to assess the entire competitive field. That is why it was particularly interesting to compare, in this context, human strategies and the actions of large language models: who would be better able to assess the level of their opponents and choose the most successful strategy against them,’ noted Dmitry Dagaev, Head of the HSE Laboratory of Sports Studies.

The experiment consisted of three tournaments: the first involved only human participants; in the second, solutions generated by popular language models were added; in the third, the number of human and machine strategies was equal. The games were played in a round-robin format: within each tournament, all strategies competed against one another once, after which they were ranked according to the number of victories. In each of the three tournaments, ten winners were selected, and a further ten prize places were awarded to the best performers based on their overall results across all three stages.

According to the results of the competition, humans proved stronger than machines. In Tournament No. 3, where the number of strategies proposed by humans and language models was equal, the best AI result was 18^th place.

‘We believe that the reason for the human victory over the machine lies in the fact that people used strategies reflecting deeper iterative reasoning compared with the more primitive strategies of the language models. It is also worth noting that in such tasks, when making decisions, players usually rely either on the rules of the game—that is, they try to select a strategy that will perform well under the given rules—or on the character of their opponent, attempting to anticipate their actions. In our case, people did not significantly change their strategies between tournaments; in other words, they were guided more by the rules themselves than by who they were playing against—whether human or AI,’ noted Dmitry Dagaev.

Petr Parshakov

‘The tournament showed that in the absence of a single clearly correct strategy, as in “The Election Race,” people were better able to predict the behaviour of other participants and propose solutions capable of outperforming AI. In the future, based on such experiments, we plan to create a benchmark for language models in order to rank them according to how closely their behaviour resembles that of humans,’ said Petr Parshakov, Head of the International Laboratory of Intangible-driven Economy.

The tournament was held as part of a grant from the Russian Science Foundation (No. 25-18-00539), entitled ‘Comparative Analysis of Behaviour of Artificial Intelligence-Based Agents and Real Individuals in Economic Decision-Making’.

 

The International Laboratory of Intangible-driven Economy (HSE University in Perm) conducts empirical research in the fields of intellectual capital, behavioural economics, technological transformation of companies, and the use of resources by small and medium-sized enterprises under conditions of uncertainty.

The Laboratory of Sports Studies at the HSE Faculty of Economic Sciences carries out theoretical research and data analysis in the sports industry. Its main areas of work include identifying optimal strategies for players and teams during matches, predicting the outcome of matches and competitions, and developing tournament rules, among others.