A hierarchical approach to designing approximate reasoning-based controllers for dynamic physical systems

Hamid R. Berenji, Yung-Yaw Chen, Chuen-Chien Lee, Jyh-Shing Roger Jang, S. MurugesanUAI 1990

TL;DR

This paper introduces a novel hierarchical approach for designing approximate reasoning-based controllers, particularly for dynamic physical systems with multiple, interacting goals. By prioritizing goals and building control rules incrementally, this method offers a more structured and potentially simpler way to design controllers compared to traditional analytical methods, especially when precise system models are unavailable.

What Problem Does This Paper Address?

Designing effective controllers for complex, dynamic physical systems is a significant challenge. Traditional analytical control methods often rely on precise mathematical models of the system, which can become intractably complex or simply infeasible to obtain for non-linear systems with many interacting components. Human experts, on the other hand, excel at controlling such systems using approximate reasoning, drawing on intuition and experience. This paper aims to bridge this gap by developing a systematic method for creating controllers that mimic this human-like approximate reasoning, enabling them to handle imprecision and achieve multiple, potentially conflicting, objectives.

Methods

The core of the paper is a new method for designing approximate reasoning-based controllers, which leverages a hierarchical structure. The process involves:

  1. Goal Identification and Prioritization: Defining a set of goals ($G$) and assigning priorities ($p(G)$) to them. This allows the controller to focus on the most critical goal first.
  2. Input and Output Parameter Definition: Identifying the input control parameters ($U$) and output linguistic values ($C$) relevant to each goal.
  3. Hierarchical Rule Acquisition:
    • Highest Priority Goal: Acquiring a set of approximate control rules ($R_1$) directly related to the highest priority goal. These rules use linguistic variables to describe system states and desired actions.
  4. Lower Priority Goals: For subsequent goals, rules are formed that incorporate the condition of the previous goal being approximately achieved. This is done by adding more preconditions to the rules, effectively creating a layered control strategy. For instance, a rule for moving a cart might also require the pole to be "almost balanced."
  5. Conflict Resolution: The paper mentions using a Max-Min compositional rule of inference for conflict resolution when multiple rules are triggered simultaneously.
  6. Defuzzification: The Center Of Area (COA) method is used to translate the fuzzy output into a crisp control action.

The authors implement this method using a rule-based computer program and a prototype hardware system to solve the cart-pole balancing problem.

Datasets

The paper does not explicitly mention using pre-existing datasets for training or evaluation in the traditional machine learning sense. Instead, the "dataset" is the dynamic physical system itself (the cart-pole balancing problem). The evaluation is performed using:

  • Computer Simulations: The controllers are tested in simulation using a set of 7 different cart-pole configurations (varying pole lengths and weights).
  • Hardware Implementation: A physical cart-pole system is built and controlled using both the proposed fuzzy controller and a conventional State Feedback Controller (SFC).

Key Insights and Ideas

  • Hierarchical Control for Interacting Goals: The most significant insight is the proposed hierarchical approach to managing multiple, interacting goals. By explicitly defining goal priorities and building rules that ensure previous goals are "approximately" met, the method provides a structured way to tackle complex control tasks that go beyond single-objective optimization.
  • Leveraging Approximate Reasoning for Robustness: The paper highlights that approximate reasoning, by using linguistic variables and fuzzy logic, can inherently handle imprecision and uncertainty in system dynamics. This makes it a strong candidate for systems where precise mathematical modeling is difficult or impossible.
  • Complementary to Analytical Control: The authors position their approach not as a replacement for, but as a complement to, conventional analytical control methods. It offers an alternative when analytical models are a bottleneck.
  • "Approximate Achievement" as a Control Strategy: The concept of "approximately achieving" a goal is key. It allows for graceful degradation of performance on a lower-priority goal when a higher-priority goal demands immediate attention, and then a smooth transition back.

Key Results

The paper presents both simulation and hardware-based results comparing their fuzzy controller (FC) with a State Feedback Controller (SFC).

  • Simulation Results (Table 1 & Figure 2):
    • The fuzzy controller generally outperformed the SFC in terms of overshoot and settling time for pole angle control. For example, for Pole-1, the FC had a max $\theta$ overshoot of 0.33 degrees compared to SFC's 1.00 degree.
  • For cart position control, the SFC sometimes achieved faster settling times, but the FC exhibited less overshoot.
  • Figure 3 shows that the FC demonstrated a more integrated interaction between the two control goals (pole balance and cart position) compared to the SFC.
  • Hardware Results (Figures 4 & 5):
    • The hardware implementation showed that the fuzzy logic controller was able to balance the pole effectively from the start.
  • While the SFC also performed well, it required a significant effort in system modeling and identification beforehand, which is a drawback for many real-world problems.
  • Robustness: The FC proved more robust to variations in pole length and weight. In one extreme case (Pole-7, the heaviest and longest pole), the SFC failed to balance the pole, while the FC managed to do so, albeit with difficulty.
  • Design Complexity: The fuzzy controller design was considered less complex as it did not require a complete system model, unlike the SFC.

Limitations and Future Work

  • Controller Modification: The authors note that fine-tuning the fuzzy logic controller might be "a little more difficult" compared to an SFC, implying more parameters to adjust.
  • Future Work: The authors suggest applying this hierarchical approach to other domains and mention plans for a larger-scale project involving the rendezvous and docking operation of the Space Shuttle with the Space Station or a satellite. They also indicate that further tests are needed to validate the approach across various applications.

Summary updated

BibTeX
@inproceedings{berenji1990hierarchical,
  author = {Hamid R. Berenji and Yung-Yaw Chen and Chuen-Chien Lee and Jyh-Shing Roger Jang and S. Murugesan},
  title = {{A hierarchical approach to designing approximate reasoning-based controllers for dynamic physical systems}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1990},
}
Figure