Making Judgments

Hans J. BerlinerAAAI 1980

Abstract

Hans Berliner's 1980 paper challenges the dominance of discrete logic in AI, proposing the SNAC method to handle "judgment" in complex domains like backgammon. By using smooth, non-linear evaluation functions, Berliner's BKG 9.8 program achieved a historic victory over a world champion, demonstrating the power of continuous mathematical modeling over rigid rule-based systems.

The Problem with Discrete Logic

In the early days of AI, problem-solving was synonymous with "reasoning": the use of discrete entities and logical branching (if-then statements). Berliner argued that this paradigm fails in "judgment domains" such as Chess or Backgammon. These environments are too vast to describe with perfect accuracy, and small environmental shifts require "graded responses" rather than binary decisions.

Traditional logic-based systems face three major hurdles:

  1. Contradictory Premises: Simple rules break down when multiple conflicting factors are introduced.
  2. The Blemish Effect: Prematurely "quantizing" continuous variables leads to search anomalies where programs make poor decisions based on arbitrary thresholds.
  3. Brittleness and Complexity: As domains grow, adding more hand-coded rules becomes computationally expensive and often leads to unintended interference between old and new rules.

The SNAC Method

To overcome these hurdles, Berliner proposed the SNAC method (Smooth, Non-linear, Application Coefficients) for constructing arithmetic evaluation functions. Mathematically, this allows a program to "shape" its search space.

  • Smoothness: Feature values vary continuously rather than jumping between discrete states.
  • Non-linearity: Terms in the polynomial are non-linear, which Berliner identifies as crucial for high-level expertise.
  • Application Coefficients: These "slowly varying variables" determine the weight of a specific feature in a given context.
Illustration of the SNAC method's ability to 'round the hills' of a search space by adjusting application coefficients to avoid local maxima.

One of the most elegant applications of SNAC is "rounding the hills." By using application coefficients to flatten the "hill" of a goal as the program nears the summit, the system can avoid getting stuck on sub-optimal local maxima and naturally transition to the next set of goals.

The Historic Victory of BKG 9.8

The effectiveness of the SNAC approach was demonstrated through the performance of the BKG 9.8 backgammon program. In 1979, BKG 9.8 achieved a historic victory over the reigning World Champion, Luigi Villa. This marked the first time a computer program defeated a world champion in a recognized intellectual game.

The critical backgammon position where BKG 9.8 chose a counter-intuitive move (breaking its own blockade) to defeat World Champion Luigi Villa.

A standout moment in the match involved a move where the program chose to break its own defensive blockade (move 13-8, 3-2). Human experts initially missed this move because they were biased toward maintaining a painstakingly built blockade. However, the SNAC-based program calculated that breaking the blockade improved its "timing" and "containment" of the opponent. This counter-intuitive judgment allowed the program to win the match.

Conclusion and Future Work

While SNAC proved highly effective, Berliner acknowledged that the system remains a "black box" of polynomials, making it difficult to determine if it mirrors human thinking. Future work suggested two paths: developing self-explanation capabilities so the program can explain its judgments in human-readable terms, and failure identification to automatically tune evaluation functions when mistakes occur.

Summary updated

BibTeX
@inproceedings{berliner1980making,
  author = {Hans J. Berliner},
  title = {{Making Judgments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure