In the early 1980s, the field of Artificial Intelligence was dominated by "Expert Systems." A seminal paper presented at the first AAAI conference in 1980 by Leonard Friedman introduced the Plausible Inference (PI) system. This research sought to move beyond the rigid constraints of binary logic to automate the troubleshooting of complex physical systems, such as spacecraft, by mimicking the human process of "sharing suspicion."
The Challenge of Diagnostic Reasoning
Traditional deductive logic typically moves from cause to effect: if $A$ implies $B$, and $A$ is true, then $B$ must be true. However, diagnostic reasoning often requires working backward. When a system failure is observed (i.e., $B$ is false), the goal is to identify which component of the antecedent $A$ failed.
Humans excel at this "process of elimination," but early AI systems struggled with three primary hurdles:
- Uncertainty: Failures are rarely 100% certain.
- Bidirectional Reasoning: The need to move from observed effects back to potential causes.
- Non-monotonicity: The ability to revise beliefs when new evidence (like a test result) emerges.
The PI System Architecture
The PI system, built upon the ERIS framework, addresses these challenges by combining formal propositional logic with a "credibility" calculus.
Credibility Scores and Augmented Logic
Every assertion in the PI system is assigned a truth value and a "credibility" score ranging from -1 (certainly false) to +1 (certainly true). This approach was inspired by the MYCIN medical expert system. To model physical processes more accurately, the system uses an augmented propositional calculus that includes a specialized "CAUSES" connective alongside standard AND/OR operators.
New Rules of Inference
Beyond standard Modus Ponens, the PI system introduces two critical mechanisms:
- Method of Confirmation: Propagates belief forward from causes to effects.
- Method of Denial: Propagates suspicion backward from failed goals to potential causes.
Case Study: Spacecraft Signal Transmission
The paper evaluates the PI system using a model of a spacecraft's communication system. This model represents the logical dependencies required for a signal to be received on Earth.
As shown in the figure above, the goal "SIGNAL RECEIVED" depends on three main branches: "GROUND ANTENNA RECEIVING," "POINT S/C" (spacecraft pointing), and "TRANSMIT SIGNAL." Each branch is assigned a transfer factor or credibility score (e.g., -0.7 for pointing and -0.3 for transmission) that reflects the a priori likelihood of failure.
The Process of Elimination
When a failure occurs—for instance, "Signal Not Received"—the system initiates a "sharing of suspicion." It marks all preconditions as "Possibly-False" (PF).
The power of the PI system lies in its ability to "vector in" on a problem through evidence:
- Initial Suspicion: If the signal is missing, all three branches (Antenna, Pointing, Transmission) are suspects.
- Evidence Integration: If a human controller performs a "roll test" confirming the spacecraft is indeed transmitting, the system uses Modus Ponens to flip the status of the "TRANSMIT SIGNAL" branch from PF to "True."
- Belief Revision: Through dependency-directed backtracking, the system automatically clears the transmission branch and increases the "credibility of failure" for the remaining suspects, such as the pointing sensors.
Limitations and Legacy
While pioneering, the PI system had notable limitations. It could not reason about "vanished" or transient events that leave no lasting state changes. Furthermore, the diagnosis was strictly limited by the granularity of the model; if a component wasn't explicitly in the semantic net, the system could not identify it as the cause.
Despite these constraints, the PI system represented a significant step toward creating AI that can handle the nuances of real-world uncertainty and the iterative nature of human troubleshooting.
Summary updated