The Promise of Practical AI : Expert Systems in the 1980s

The Promise of Practical AI : Expert Systems in the 1980s

In the 1980s, research into human-like artificial general intelligence was stalling due to the immense challenges of replicating the breadth of human cognition. However, a more practical application of AI emerged – expert systems. Expert systems were computer programs designed to emulate and automate the decision-making capabilities of human specialists in narrow domains.

While artificial general intelligence sought to mimic the full range of human intellectual faculties, expert systems focused on specialized skills and expertise. By encoding the logical rules and knowledge distilled from human experts, expert systems could perform tasks like medical diagnosis, mineral exploration, financial planning, and more.

This demonstrated the possibilities for artificial intelligence to have meaningful real-world impact even without achieving the lofty goal of human-level AI.

Knowledge Engineering

The core component of an expert system was its knowledge base – the repository containing the facts, concepts, logical rules, and procedures representing human expertise in a field. Knowledge engineers worked intensively with human domain experts to methodically elicit this knowledge, often through extensive interviews and observation of their reasoning across test cases.

For example, knowledge engineers creating a medical diagnostic system would work closely with physician specialists, probing their thought process during diagnosis of actual patients. The doctors’ knowledge was then precisely translated into a set of production rules with IF-THEN structure. An example rule could be:

IF a patient has symptoms A, B and C
AND test X shows abnormality D
THEN the likely diagnosis is disease Z

This knowledge engineering process was challenging and tedious. Human experts often struggled to articulate their implicit thought patterns and tacit knowledge. Knowledge engineers had to deeply probe the nuances and caveats around the rules to capture sufficient subtleties.

Despite the difficulties, building a comprehensive knowledge base covering the scope of the domain was fundamental for enabling the system’s reasoning capabilities.

Some expert system tools like EMYCIN provided a generic inference engine while the knowledge base was customized for the specific domain. This simplified development, allowing focus on the knowledge itself.

However, the inference engine imposed limitations on how knowledge could be represented. Building the domain-specific knowledge base from scratch required extensive labor-intensive analysis of human expertise.

Influential Medical Diagnosis Systems

One of the earliest and most influential expert systems was MYCIN, developed at Stanford in the 1970s to diagnose bacterial infections. MYCIN analyzed a patient’s symptoms and incrementally homed in on a diagnosis by asking a series of questions.

Once the bacterial infection was identified, it prescribed suitable antibiotics, specifying dosage levels tailored for factors like the patient’s age and kidney function.

In rigorous assessments, MYCIN demonstrated diagnostic performance comparable to infectious disease experts. This provided exciting proof that expert systems could successfully replicate specialized human skills and judgement, at least within a defined domain. However, despite its capabilities, MYCIN never saw real clinical use due to regulatory hurdles regarding AI diagnosing real patients.

Other medical diagnosis expert systems like PUFF and Internist-I followed, assisting doctors in specialty areas like lung disorders and internal medicine. These systems served as diagnostic assistants, providing second opinions and filling availability gaps for human experts. While not replacing physicians, expert systems could complement them, enhancing diagnostic quality and accessibility.

Scientific Discovery and Analysis

DENDRAL, one of the earliest expert systems created at Stanford in the 1960s, vividly demonstrated the potential of expert systems in science and analytic tasks. DENDRAL analyzed chemical spectroscopy data to deduce the molecular structure of organic chemical compounds.

It did this by deductively eliminating structural possibilities based on encoded rules of chemistry and mass spectrometry principles, along with experimental evidence from the input data. DENDRAL proved substantially superior at determining molecular structures compared to trained human chemists.

Similarly, the PROSPECTOR system helped geologists locate minerals and valuable ores by recommending geological sites likely to yield discoveries based on terrain conditions and existing drill sample data. It replicated the reasoning of an expert prospector with years of experience.

Another system called EXPLORER identified promising new oil drilling locations by matching a range of geological factors to patterns from past successful drilling sites. Expert systems proved invaluable in replicating the complex analysis required for scientific discovery and evaluation.

Optimizing Computer Systems

Digital Equipment Corporation developed an expert system called XCON to advise technicians on properly configuring components and peripherals when building VAX minicomputers per customer specifications.

The system contained expansive knowledge on VAX hardware options, performance needs, software compatibility issues, and more. By quickly applying this expertise, XCON avoided misconfigurations which previously led to post-sale fixes and rework. It was estimated to save Digital Equipment around $40 million annually in the 1980s.

The huge success of XCON highlighted business applications of expert systems in manufacturing, sales, process control, and other areas. AI could optimize and improve business efficiency without entirely eliminating human roles.

For example, Blue Cross partnered with Teknowledge to develop an expert system to facilitate speedy health insurance policy approval. United Airlines worked with Carnegie Group on an agent advisor system for recommending pricing adjustments for ticket purchases.

Automated Trading and Financial Reasoning

Expert systems also enabled automated high-level financial decision making abilities for stock trading, portfolio management, and other investment activities. Companies like Teknowledge and Intellicorp created tools applying technical trading indicators, performing due diligence on potential investments, and managing portfolio risk exposure. Other systems from KEE and KnowledgeCraft focused on business planning and financial analysis applications.

For example, a system called MoneySense served as an expert assistant for commodities traders. It encoded knowledge on market conditions, pricing trends, and successful trading strategies.

MoneySense could detect arbitrage opportunities, warn of excessively risky investments, and even educate novice traders on effective practices. Expert systems applied the distilled expertise of successful veteran traders to achieve investing optimization exceeding human capability.

While most beneficial in specialized subfields, financial expert systems found niches automating complex analytical skills and arcane knowledge. They demonstrated AI’s potential for significant real-world impact without achieving artificial general intelligence.

The Scalability Bottleneck

Despite exhibiting intelligence in narrow tasks, expert systems lacked basic common sense reasoning abilities. Their reasoning was restricted to a single specialized domain defined by their encoded knowledge base.

Expanding to new domains or broadening capabilities required massive additional knowledge engineering efforts. This knowledge acquisition bottleneck severely hampered mainstream adoption of expert systems as a general technology.

The professional labor and knowledge extraction costs made building new expert systems expensive and time-consuming. The knowledge limitations also risked brittleness and errors when faced with atypical edge cases. Still, by replicating and productizing specialized human expertise, expert systems demonstrated AI’s potential, even if scaling the approach remained challenging.

Glimpsing the Future

The pioneering expert systems of the 1980s offered a glimpse into the possibilities of applied AI. Though not as futuristic as artificial general intelligence, expert systems showed that encoding human knowledge into programs could replicate expertise, optimize processes, and enhance decision-making. This spurred proliferation of expert systems and knowledge-based applications, though inflated expectations were curbed by the challenges of knowledge engineering.

The legacy of expert systems continues today in many specialized AI programs built on expert-defined rules and heuristics. From medical diagnosis to financial analysis, expert systems established that artificial intelligence technologies could transition from pure research into practical real-world solutions.

Though narrow in scope, expert systems fulfilled the promise of augmenting human capabilities with AI, even if not matching the flexibility of human cognition. The expert systems field provided an early glimpse of the versatile future landscape for applied artificial intelligence.