Core Concepts
This paper proposes and evaluates a system that expands a gradient descent-trained defensible expert system to be a defensible Blackboard Architecture, adding activation functions and a best path-based optimization algorithm.
Abstract
The paper presents a system that combines the capabilities of expert systems and neural networks. It builds upon previous work on "defensible" systems that use gradient descent to optimize a rule-fact network, preventing the learning of spurious correlations.
The key aspects of the proposed system are:
It extends the defensible expert system to a Blackboard Architecture, adding the capability to actuate the environment through actions.
It incorporates activation functions, similar to those used in neural networks, to regulate the range of possible output values.
It proposes a best path-based optimization algorithm to train the system using gradient descent.
The paper evaluates the performance of this system through extensive simulations, analyzing the impact of various parameters such as network size, training iterations, training velocity, and trigger thresholds. Key findings include:
Increasing the number of rules in the network reduces the error, but with diminishing returns.
Training the system can improve performance, but care must be taken to avoid over-training. Selecting the best result from the training process is important.
Increasing the training velocity consistently improves performance.
Applying trigger thresholds to restrict the range of output values can reduce error, even without application-specific tuning.
The runtime of the system does not correlate well with its error level, so runtime cannot be used as a surrogate metric for accuracy.
Overall, the results demonstrate the potential of the proposed gradient descent-trained defensible Blackboard Architecture system and provide insights into its performance characteristics.
Stats
The system's average runtime increases at a greater-than-linear rate as the number of rules in the network increases.
As the number of facts in the network increases, the average error first increases and then slightly decreases.
Increasing the number of rules in the network decreases the average error, with diminishing returns.
Increasing the number of actions in the network generally increases the average error.
Quotes
"A variety of forms of artificial intelligence systems have been developed. Two well-known techniques are neural networks and rule-fact expert systems."
"Due to the growing prevalence of artificial intelligence systems, concerns have been raised that many techniques – and neural networks, in particular – lack understandability and accountability."
"This defensible system was based on an expert system rule-fact network and, thus, lacked actualization capabilities. It also did not include the activation functions used by neural networks."