Introduction
In the early 1980s, artificial intelligence planning was largely confined to "blocks world" scenarios—environments where every action had a predictable, certain outcome. However, high-stakes domains like oncology present a far more complex challenge. In cancer therapy, goals are often ill-defined, balancing the desire to cure a tumor against the risk of lethal toxicity. Furthermore, the effects of drugs are highly variable across different patients.
ONYX was developed to move beyond the rigid "if-then" protocols of systems like ONCOCIN. Instead of relying on compiled knowledge that breaks down in non-standard cases, ONYX seeks to reason from "first principles" using a mechanistic model of the human body to navigate medical uncertainty.
The ONYX Architecture
The ONYX system employs a three-stage pipeline to handle planning in ill-structured domains:
- Plan Generation: Rather than searching an exhaustive state space, ONYX uses a therapy hierarchy and production rules to identify a small set of plausible plans. Control rules navigate this hierarchy, while generation rules propose specific actions—such as dose reductions—based on individual patient data.
- Symbolic Simulation: This stage serves as the core innovation of the architecture. Recognizing that precise quantitative models of human physiology are often too complex or unavailable, ONYX uses qualitative rules to predict how drugs affect cell populations over time.
- Plan Ranking: Once simulations are complete, the system uses decision analysis to rank plans based on three primary objectives: improving prognosis, minimizing treatment risk, and adhering to established protocols.
Symbolic Simulation and Hierarchical Modeling
To make simulation computationally tractable, ONYX represents the human body as a hierarchy of interconnected, communicating objects. For instance, the "body" model contains a "bone marrow" model, which in turn contains a "marrow space" model. These models use message passing to simulate complex interactions, such as the movement of a drug from the bloodstream into the bone marrow.
This approach allows for qualitative reasoning. In medicine, clinicians often know the direction of a change (e.g., a drug will decrease white blood cell counts) even if the exact magnitude is unknown. Symbolic simulation allows the planner to reason about these trends without requiring perfect differential equations.
Results and Visual Feedback
The ONYX prototype, implemented on a Xerox 1108 LISP workstation, successfully demonstrated the ability to simulate physiological responses to chemotherapy. By drawing on the knowledge base of the ONCOCIN system, the authors were able to model bone marrow responses to hypothetical anti-cancer drugs.
One of the most significant outputs of the system was its ability to provide visual feedback. By generating graphical plots of predicted cell depletion and recovery, the system offered a "visual explanation" for its recommendations, helping clinicians understand why a specific plan might be considered too risky.
Discussion and Future Work
While ONYX proved the viability of a "generate-simulate-rank" architecture, several challenges remained at the time of its publication. The three components were not yet fully integrated into an automated loop, and the simulations were largely deterministic. The authors noted that future iterations would need to incorporate probabilistic outcomes to better represent the true uncertainty of medical treatment. Additionally, the high cost of knowledge engineering required to build detailed mechanistic models for every organ system remains a significant hurdle for scalability.
Summary updated