ONYX: An Architecture for Planning in Uncertain Environments

TL;DR

This paper introduces ONYX, a novel AI planning architecture designed for complex domains like cancer therapy where goals are fuzzy and actions have unpredictable outcomes. It tackles this by combining strategic planning with mechanistic simulation, allowing it to generate, predict, and rank plausible treatment plans even when traditional methods fail.

What Problem Does This Paper Address?

Traditional AI planning systems, like STRIPS, excel when goals are clear and actions have deterministic effects. They operate by searching through a "state space" using predefined operators. However, many real-world problems, particularly in medicine, defy this neat structure. In cancer therapy, for instance, treatment goals are often ill-defined (e.g., "improve prognosis" rather than a specific numerical target), and the effects of drugs or radiation are uncertain and can lead to unforeseen side effects or interactions. Existing systems like ONCOCIN rely on compiled knowledge (protocols and algorithms) which work well for standard cases but falter when faced with novel situations or complex patient-specific issues. ONYX aims to bridge this gap by providing expert-quality therapy suggestions in these difficult, uncertain scenarios.

Methods

ONYX employs a three-step planning process:

  1. Plan Generation: This phase uses a hierarchical representation of treatment strategies. Control rules guide the exploration of this hierarchy, deciding which branches are plausible given the patient's current state. Generation rules then propose specific plan steps (e.g., adjusting drug dosage) based on patient data and domain knowledge. The goal is to create a small set of "plausible plans" that are consistent with the patient's current condition.
  2. Symbolic Simulation: Unlike traditional deterministic models, ONYX uses a qualitative, symbolic simulation approach. It models the domain (e.g., the human body) hierarchically, with each component having its own behavior rules. This allows it to predict the future states of the patient after executing each plausible plan. The simulation focuses on qualitative trends and relationships rather than precise numerical predictions, acknowledging the inherent uncertainty.
  3. Plan Ranking: This step, though still under development, aims to rank the generated plans based on how well their predicted outcomes meet the therapy goals. This will involve decision analytic techniques, incorporating probabilities of outcomes (from the simulation) and patient preferences (e.g., improving prognosis, reducing risk, adhering to protocols).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Bridging Strategic and Mechanistic Knowledge: ONYX's core innovation is its integration of high-level strategic planning (how to approach therapy) with mechanistic simulation (how the body actually responds). This allows for more robust planning in uncertain domains.
  • Qualitative Simulation for Uncertainty: Moving away from deterministic mathematical models, ONYX's symbolic simulation offers a more realistic way to predict outcomes in complex biological systems where precise quantification is often impossible.
  • Hierarchical Planning Representation: Using a therapy hierarchy for plan generation allows for a structured exploration of treatment options, starting from broad choices and drilling down to specific actions.
  • Handling Ill-Specified Goals: The plan ranking phase, by incorporating patient preferences and goals, aims to address the challenge of planning when objectives are not precisely defined.

Key Results

The paper describes the architecture and the conceptual framework of ONYX. While the plan ranking phase is noted as substantially unimplemented, the plan generation and symbolic simulation components are described with examples. The symbolic simulation can predict the behavior of the bone marrow in response to an anti-cancer drug, showing how cell populations change over time (illustrated by a plot in Figure 4). This provides a basis for estimating how well a plan might meet treatment goals.

Limitations and Future Work

The authors explicitly state that the plan ranking phase is "substantially unimplemented." They plan to augment the symbolic simulation to provide probabilities of outcomes and to integrate patient preferences into the ranking process. A key future direction is to combine the three components of ONYX into a coherent planning system and to integrate ONYX's strategic and mechanistic knowledge with the protocol-based knowledge from ONCOCIN to enhance its decision-making and explanatory capabilities.

Summary updated

Abstract

The ONYX program is designed to fill the need for planning in application areas where traditional planning methodology is difficult to apply. While the program being developed will assist with the planning of cancer therapy, its architecture is intended to be of use whenever goals are ill-specified, plan operators have uncertain effects, or trade-offs and unresolvable conflicts occur between goals. We describe a planning process which uses strategic information and a mechanistic model of the domain. The process consists of three steps: (1) generate a small set of plausible plans based on current data, (2) simulate those plans to predict their possible consequences, and (3) based on the results of those simulations, rank the plans according to how well each meets the goals for the situation.
BibTeX
@inproceedings{1985onyx,
  title = {{ONYX: An Architecture for Planning in Uncertain Environments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure