Jacques Pitrat, l'Intelligence Artificielle et les Jeux

Tristan Cazenave

Introduction

Jacques Pitrat was a visionary pioneer who envisioned "General Game Playing" (GGP) decades before it became a mainstream AI milestone. This paper provides a retrospective on Pitrat’s career, tracing the lineage of his ideas from symbolic learning and "meta-intelligence" to the bootstrapping techniques that underpin modern champions like AlphaZero and Polygames.

The Quest for Generality

For much of the 20th century, AI research in games focused on "ultra-specialization"—building machines like Deep Blue that were world-class at Chess but incapable of playing a simple game of Tic-Tac-Toe. Jacques Pitrat saw this as a dead end for true artificial intelligence.

Pitrat’s work addressed the challenge of generality: how to build a single system capable of learning and playing any game given only its rules. He advocated for moving away from brute-force search toward "parsimonious" learning, where a system understands why a move is good rather than just memorizing millions of data points.

Methodologies of Meta-Intelligence

Pitrat championed several key methodologies that diverged from the purely statistical approaches common today:

  • General Game Playing (GGP): In 1968, Pitrat published the first paper on a general game system. These systems use a formal Game Description Language (GDL) based on predicate logic, allowing an agent to parse rules and develop strategies for games it has never seen before.
  • Explanation-Based Generalization (EBG): Unlike modern Deep Learning, which requires massive datasets, EBG is a symbolic approach. It uses game rules to reason about successful move combinations, generalizing specific instances into first-order logic rules.
  • Bootstrapping (Amorçage): Pitrat was obsessed with "meta-intelligence." He developed systems like CAIA (an Artificial Researcher in AI) designed to write its own programs and improve its own search algorithms.
  • Nested Monte Carlo Tree Search (MCTS): While Pitrat was skeptical of the randomness in early MCTS, his bootstrapping philosophy led to the development of Nested MCTS. In this approach, a search process uses lower-level simulations to "prime" or bootstrap its own performance.
Figure 4: Visual representation of search levels and bootstrapping in game trees, illustrating the progression from simple simulations to nested search structures.

Benchmarks and Experimental Validation

As a theoretical and algorithmic retrospective, Pitrat's work utilized a variety of game environments as benchmarks over 50 years. These included classic games like Chess and Go, as well as GGP benchmarks like Hex (high branching factor) and Breakthrough (low branching factor). Experimental games like Phutball (Philosopher's Football) were used in "Metaknowledge" colloquiums to test how different AI architectures—specifically policy-based versus evaluation-based—performed against each other.

Key Results and Legacy

Pitrat’s principles have seen significant modern validation:

  • Competition Success: Programs built on his principles, such as ARY, won major GGP competitions at IJCAI 2009 and AAAI 2010.
  • The Power of Evaluation: The program Athénan, which uses Minimax search and learned evaluation functions (eschewing MCTS), won 11 gold medals at the 2021 Computer Olympiad. This proved that Pitrat’s focus on refined, parsimonious evaluation remains highly competitive.
  • Foundational Impact: Pitrat supervised numerous PhD theses that laid the groundwork for the "MCTS Revolution" in the mid-2000s, which eventually led to the defeat of the world champion in Go.

Conclusion

Jacques Pitrat remained critical of the statistical "black box" nature of modern MCTS and Deep Learning, preferring models of "reflexive consciousness" and rule-based systems. The future of the field likely lies in reconciling these two worlds: combining the symbolic generalization of Pitrat’s era—which allows for safe, explainable, and data-efficient learning—with the statistical power of modern neural networks.

Summary updated

BibTeX
@article{cazenave2022jacques,
  author = {Tristan Cazenave},
  title = {{Jacques Pitrat, l'Intelligence Artificielle et les Jeux}},
  journal = {Rev. Ouverte Intell. Artif.},
  year = {2022},
  doi = {10.5802/roia.22},
}
Figure