EXPLORA: AI/ML EXPLAINABILITY FOR THE OPEN RAN
With the foreseen rapidly changing environment approaching with the sixth generation (6G), the Open Radio Access Network (RAN) paradigm has transformed cellular networks into visualised, software-based components. EXPLORA is a new framework that helps to explain DRL-based solutions for cellular networks through a clear and concise manner in comparison to solely providing Deep Reinforcement Learning (DRL)-based solutions which are more difficult to explain in the context of resource allocation problems.
A publication by Claudio Fiandrino, Leonardo Bonati, Salvatore D’Oro, Michele Polese, Tommaso Melodia, and Jeorg Widmer
EXPLORA: A Simplified Framework
EXPLORA makes the explanation of DRL-based control solutions for the Open RAN ecosystem by synthesising network-oriented that are based on an attributed graph that produces a connection between a DRL agent and input state space. By doing this, EXPLORA can explain the models by providing information on the wireless context that the DRL agent operates on.
Through the framework, the outcome of the explanations can be used for performing informative action steering and median transmission irate improvements of 4% and tail improvements of 10%. The agent’s behaviour is then determined by the effect of its decisions on the environment from details of the contribution of each component of the multi-modal action.
How It Works
DRL agents leverage deep neural networks which can be quite complicated to not only explain but also understand. Therefore, DRLs are difficult to use in production networks as there is a lack of understanding of the logic behind decisions. However, DRL is not the only factor that is affected as AI is also affected by this issue as these models tend to be harder to understand. As a result, the authors try to bridge the gap and make DRL for Open RAN applications explainable which lead to the creation of EXPLORA.
EXPLORA can simplify knowledge by analysing transitions between actions, as it can then characterise the individual contribution of each component of a multi-modal action. The framework does not solely reveal how inner mechanisms of a model works, but also provides explanations on wireless network behaviour to help operators in interpreting AI decisions. Along with this, it makes it possible to perform informed and targeted ad-hoc adjustments to the agent’s behaviour to modify its decision-making process with the goal of improving the overall network performance.
Key features
EXPLORA consists of the XAI module and the XAI-aided Explanation-Driven Behaviour Refiner (EDBR) module which generates post-hoc explanations about the agent behavior by building the attributed graph G throughout an observation window W. These explanations are gathered to get a better understanding of the decision-making process which then can identify the inefficiencies and improve network performance overall. To distil the knowledge, EXPLORA analyses transition between actions (edges in G) to characterize the individual contribution of each component in the multi-modal action. With the generated distilled information, it makes it possible to perform informed and targeted ad-hoc adjustments to the agent’s behavior to modify its decision-making process to achieve the overall goal of improving network performance.
Trustworthy AI in Open RAN
There were three strategies used to implement EXPLORA to make it an operational component of the near-RT RIC. It has been used for a base component of the RIC itself hosted outside the xApp domain, a component of each xApp embedded in microservices that execute the DRL agent, or a standalone xApp that interacts with more xApps hosting DRL agents. EXPLORA can be used within and outside O-RAN networks by interfacing with AI/ML algorithms running on the BSs directly or on controlled networks not tied to O-RAN specifications. Along with this, EXPLORA can be utilized within production environments as it can speed up AI/ML solutions. Within production environments, they can be used to manage complex and larger networks.
Improving Overall Performance
This new framework has a variety of benefits that were tested and proven to be sustainable and successful. Overall, the main benefit found from EXPLORA has been the ability to synthesize model explanations that are used for monitoring and troubleshooting for domain experts. Within an Open RAN scenario and being applied to a set of DRL agents executing as O-RAN xApps that govern RAN slicing and scheduling policies, EXPLORA has given the output of clear explanations and improved overall performance to identify and substitute actions that would normally lead to low expected rewards programmatically.