Aiora Artemis, an RF engineering optimizer developed by FILPAL, utilizes reinforcement learning as the main AI framework. This article explains the potential of the 3 main classes of AI and why reinforcement learning is best used for AI-assisted optimization.
AI Trend
The general hype for AI in these recent years is being felt across the world. Since the launching of ChatGPT in November 2022, the world has gone into a frenzy and scramble for anything and everything related with AI.
One particular field of interest in AI technology is the attractive idea of Generative AI. In simple words, generative AI key strength is in its ability to generate ideas, designs and contents based on prompts.
It is simultaneously perceived as both a boon and a threat by societies around the world. Boon by sectors such as engineering that allows quick generation of designs and ideas, thus increasing design speed; Threat by sectors such as artistry and content creators who see their lifelong skill development taken over by AI almost overnight.
In short, the simple fact is this, AI is now, and it is here to stay.
Classes of AI
But Generative AI is not the only AI technology that is growing. In fact, there exist 3 common types of AI.
- Unsupervised Learning AI (Generative)
- Supervised Learning AI (Discriminative)
- Reinforcement Learning AI
Each type of AI adheres to a behavioral pattern on how it performs its training.
Generative AI Training
The keyword to generative AI is inclusivity. In training generative AI, this model evaluates datasets by identifying variables that are not yet found in its model. It then proceeds to learn the patterns and relationships of these variables with respect to the existing model and integrate it back into the model. This method of unsupervised machine learning allows the AI to continuously absorb new ideas and expand its decision boundaries.
The result is an AI that is capable of generating new and original content based on its understanding the relationship and patterns of these variables. The more, the larger the datasets fed into the model, the better the AI is able to mimic and create contents based on what it has learnt.
The Problem with Generative AI
Though impressive, generative AI main limitation lies in its inability to effectively create novel ideas outside of what it learnt. Generative AI decision boundaries lies within what is known, created or discovered by mankind.
It can generate permutations and seemingly creative content based on what it knows, but its rarely capable of effectively generate new and novel designs.
In the field of RF engineering, this means that any designs generated falls within well-established and researched methodologies and topologies, making it challenging to “break out of the box” for novel methods.
Discriminative AI Training
You can consider discriminative AI as the anti-thesis of Generative. Where generative AI seeks inclusivity, discriminative AI seeks exclusivity Discriminative AI is a supervised learning method of interpreting datasets to learn the underlying variables in differentiating classes of data. Labelled datasets are fed into the AI training to improves its ability to recognize the decision boundaries between classes.
The role of discriminative AI is found in industry that requires quick, automated and effective way of telling A from B. Consider efforts such as fruit harvesting, robotics with this AI will be able to firstly differentiate the fruits from its environment, and proceed to identify if the fruit is unripe, ripe or overripe, and process them accordingly.
The Problem with Discriminative AI
Similar to generative AI, and in some aspect even more problematic, is the reliance of discriminative AI on supervision. In essence, the decision-making ability originates from human inputs, with discriminative adhering strictly to the set boundaries of the human input. A lack of clarity in the labelling of the dataset would confuse the AI into misinterpreting inputs.
In the field of RF engineering where many design process have a certain level of ambiguity, its usage is limited due to the challenges in classifying and categorizing RF designs, limiting the AI from able to identify and suggest corrective actions.
Reinforcement Learning AI Training
To some extents, reinforcement learning is related with discriminative AI. It relies on supervised learning and its purpose also centers around decision making. In reinforcement learning, an agent performs decision making using a modelled trial-and-error methodology. Each decision is then subjected to a feedback mechanism (objective function). A good decision will be “rewarded” and a bad decision will be “penalized”. The feedback is then used to strengthen the AI policy used for modelling the trial-and-error methodology. Reinforcement learning goal is to achieve maximum cumulative reward with every pass, allowing the AI to converge on a solution.
Reinforcement Learning in RF Engineering Optimization
In RF engineering optimization process, the usage of reinforcement learning proves greatly beneficial. In the discussions above, we recognize:
- Generative AI inability to “think outside the box”
- Discriminative AI struggles with ambiguity
In RF design process, novel designs and topologies are often derived from new mathematical modelling methods, which are then synthesized into 3D models with physical dimensions. The challenges are often found in obtaining the expected performance due to
- complexity of the design
- real-world material science variables
- environmental factors
- etc
The traditional optimization process relies on optimization algorithm such as Genetic Algorithm, Particle Swarm Optimization, Gradient and much more, with their own unique objective functions. However, these algorithms often struggle with speed and accuracy when designs become more complex.
By using reinforcement learning as an optimization algorithm, the challenges associated with design complexity are addressed. Using the trial-and-error and feedback mechanism, novel and “out of the box” designs are addressed much more easily while mitigating the challenges associated with ambiguity from complexity that usually comes from very complex designs.
In fact, due to the behavior of reinforcement learning AI to reach for cumulative reward, it greatly improves the speed and accuracy of convergence in optimization.
Conclusion
In the field of RF engineering, the optimization process for complex process takes up an approximate of 30% to 70% of the time in the design process with concerns of speed, accuracy and the infamous premature convergence often faced by traditional algorithms. The usage of AI as the optimizer must address the concerns of achieving good accuracy in novel and computationally complex designs. Comparing the 3 AI framework, Reinforcement Learning stands out as the best candidate for the job.
Aiora Artemis is the latest innovation based on the Aiora AI engine, complementing existing RF EDA tools in the market for boosting RF designs optimization speed and accuracy. Find out more: here
Originally published at http://filpal.wordpress.com on April 27, 2024.
