Problem-Solving as a Language: A Computational Lens into Human and Monkey Intelligence

Qianli Yang, Zhihua Zhu, Ruoguang Si, Yunwei Li, Jiaxiang Zhang, Tianming YangbioRxiv 2024 · DOI

TL;DR

This paper introduces the "Language of Problem-Solving" (LoPS) model, a novel computational framework that treats problem-solving as a language with its own grammar. By analyzing gameplay in a modified Pac-Man game, the authors reveal striking differences in the complexity and hierarchical structure of problem-solving between humans and monkeys, suggesting an evolutionary continuum and demonstrating how this "language" evolves with learning.

What Problem Does This Paper Address?

Quantifying and comparing problem-solving abilities across species, and understanding its evolutionary origins, is a fundamental challenge in cognitive science. While we know humans are exceptional problem solvers, precisely how this ability is structured, how it differs from other species, and how it develops remains elusive. Traditional methods often rely on subjective reports, which are impossible for non-human primates, or focus on simple action sequences that miss the underlying cognitive structure. This paper aims to bridge this gap by developing a quantitative framework to extract the latent structure of problem-solving behavior.

Methods

The core of this research is the Language of Problem-Solving (LoPS) model. This framework conceptualizes problem-solving as a sequence of "words" (strategies) governed by a "grammar" (temporal organization of strategies).

  1. Cross-Species Paradigm: The classic Pac-Man game was adapted for both humans and macaque monkeys. This game provides a rich environment for complex decision-making, involving foraging, avoiding threats, and utilizing power-ups.
  2. Strategy Extraction: The authors first identified a set of heuristic strategies (e.g., "local," "global," "evade," "energizer") that could explain the observed joystick movements of both species.
  3. UpN-gram Model: A novel language model, the UpN-gram, was developed. Unlike standard n-grams that only consider sequential dependencies between words (strategies), the UpN-gram also incorporates the influence of underlying game states, capturing a more nuanced, non-Markovian temporal structure.
  4. Grammar Induction: An inverse problem approach was used to "induce" the LoPS grammar from observed gameplay data. This involves converting action sequences into strategy sequences and then using a probabilistic graphical model (specifically, a Bayesian network) to infer the grammar that best explains the observed behavior.
  5. Complexity Metrics: LoPS complexity was defined as a measure of grammar length and hierarchical depth, providing a quantifiable metric for problem-solving sophistication.

Datasets

The study utilized behavioral data collected from:

  • 34 healthy human participants.
  • 2 macaque monkeys.

The data consisted of gameplay recordings from the modified Pac-Man game, including visual inputs, joystick/keyboard actions, and game states. The monkeys' data was collected over three years, while human participants were tested in two sessions.

Key Insights and Ideas

  • Problem-Solving as Language: The most significant insight is the successful modeling of problem-solving as a structured language with its own grammar. This provides a powerful new lens for analyzing cognitive processes.
  • Hierarchical Structure: Human problem-solving exhibits a significantly more complex and hierarchical grammar compared to monkeys, suggesting a deeper level of abstraction and planning.
  • Evolutionary Continuum: Monkey and novice human grammars show remarkable similarity, suggesting a shared evolutionary root for problem-solving strategies, with humans developing more sophisticated structures over time.
  • Learning and Evolution: Both humans and monkeys demonstrate that their LoPS grammars evolve with experience, becoming more complex and efficient, indicating that problem-solving is a dynamic, learned skill.
  • State-Dependent Grammars: The UpN-gram model's ability to incorporate game states into grammar induction highlights how environmental context shapes problem-solving strategies.

Key Results

  • Human vs. Monkey Complexity: Expert human players demonstrated significantly higher LoPS complexity (average grammar length) than novice humans and monkeys (p < 0.001).
  • Performance Correlation: LoPS complexity was positively correlated with game performance (total score, r = 0.65) and negatively correlated with reaction time (r = -0.44), indicating that more complex grammars lead to better and faster problem-solving.
  • Learning Trajectories: Both monkeys and humans showed an increase in LoPS complexity over time, with humans developing more complex grammars and strategies, particularly those involving energizers.
  • Grammar Structure Differences: Markov networks of game state variables revealed that expert humans integrate ghost-related and foraging-related states via energizers, a connection largely absent in novices and monkeys.

Limitations and Future Work

  • UpN-gram Limitations: The authors acknowledge that the UpN-gram model has limitations, including its fixed history length and potential data sparsity. They suggest exploring more advanced language models like RNNs and transformers for richer primitive representation and compositional rules.
  • Neural Substrates: While the study highlights hierarchical cognitive control, future work could explore the neural underpinnings of these LoPS grammars, potentially linking them to specific brain regions like the prefrontal cortex.
  • Generalizability: The Pac-Man paradigm, while effective, is a specific domain. Future research could investigate the applicability of the LoPS framework to other complex problem-solving tasks.
  • Cross-Species Learning: Further investigation into how learning mechanisms differ between species and how they contribute to grammar evolution would be valuable.

Summary updated

Abstract

1 Abstract Human intelligence is characterized by our remarkable ability to solve complex problems. This involves planning a sequence of actions that leads us from an initial state to a desired goal state. Quantifying and comparing problem-solving capabilities across species and finding its evolutional roots is a fundamental challenge in cognitive science, and is critical for understanding how the brain carries out this intricate process. In this study, we introduce the Language of Problem-Solving (LoPS) model as a novel quantitative framework that investigates the structure of problem-solving behavior through a language model. We adapted the classic Pac-Man game as a cross-species behavioral paradigm to test both humans and macaque monkeys. Using the LoPS model, we extracted the latent structure — or grammar — embedded in the agents’ gameplay, revealing the non-Markovian temporal structure of their problem-solving behavior. The LoPS model captured fine-grained individual differences among the players and revealed the striking differences in the complexity and hierarchical organization of problem-solving behavior between humans and monkeys, reflecting the distinct cognitive capabilities of each species. Furthermore, both humans and monkeys evolved their LoPS grammars during learning, progressing from simpler to more complex ones, suggesting that the language of problem-solving is not fixed, but rather evolves to support more sophisticated and efficient problem-solving. Through the lens of a language model, our study provides insights into how humans and monkeys break down problem-solving into compositional units and navigate complex tasks. This framework deepens our understanding of human intelligence and its evolution, and establishes a foundation for future investigations of the neural mechanisms of problem-solving.
BibTeX
@misc{yang2024problem,
  author = {Qianli Yang and Zhihua Zhu and Ruoguang Si and Yunwei Li and Jiaxiang Zhang and Tianming Yang and Tianming Yang},
  title = {{Problem-Solving as a Language: A Computational Lens into Human and Monkey Intelligence}},
  year = {2024},
  doi = {10.1101/2024.04.12.589234},
  howpublished = {bioRxiv},
}
Figure