Chiou Peng Lam

Explainable AI (XAI) is becoming increasingly vital in healthcare, particularly in understanding decisions made by complex, black-box models in medical diagnostics. Amid discussions on balancing AI's predictive accuracy with interpretability, 'post-hoc' XAI methods are a potential solution, offering interpretability after model training without compromising diagnostic precision. Counterfactual explanations are one such 'post-hoc' XAI technique, where these explanations modify input data, like features in an X-ray image, to see how such changes affect AI image-driven diagnoses, offering insights into the model's reasoning. However, the complexity of medical image data poses challenges in generating these counterfactuals. To address this challenge, we propose a novel Cooperative Co-evolution approach with dynamic decomposition for explainability in AI for medical diagnostics by evolving visual counterfactual explanations for medical images without requiring preliminary assumptions about image decomposition. The method, evaluated using MNIST and Diabetic Foot Osteomyelitis datasets and benchmarked against a genetic algorithm, was capable of generating counterfactual explanations for medical images, showing advantages with larger images.

In cooperative co-evolutionary approaches, a larger global problem is typically decomposed into multiple sub-components, with solutions concurrently evolved for each. Problem decomposition commonly occurs prior to the execution of the algorithm, but this assumes that the optimal decomposition is already known. However, the optimal decomposition is typically unknown for problems with a large number of components, such as multi-agent systems featuring many agents, where the challenge of decomposing the problem grows with the scale of the number of agents involved. In this paper, we propose a dynamic decomposition scheme based on determining optimal groupings of agents, whilst simultaneously evolving the agent behaviour, specifically addressing scalability challenges inherent in multi-agent systems featuring a large number of agents. The approach was evaluated in a real-time strategy game by evolving the behaviour of a heterogeneous team of agents representing combat units. Four scenarios were investigated featuring an increasingly larger number of agents. We compared our approach against two cooperative co-evolutionary algorithms involving commonly used static decomposition schemes. The proposed dynamic decomposition approach matched or exceeded the performance of these two static schemes across all scenarios.

Complex adversarial operations typically involve the allocation of finite resources to meet a set of objectives over a number of phases. This poses a challenge for AI-based strategy discovery. A strategy for one phase cannot be developed in isolation as the resources available in any one phase are dependent on the outcome of previous phases. Our proposed solution is to combine an evolutionary algorithm search with human-guided evaluation. The approach uses simulation-based fitness evaluation, where a human operator can view the fittest solution after every set number of generations. The operator can 'lock in' strategies for particular phases, and 'suggest' alternative strategies to guide further evolution. Key to our approach is a representation encoding that allows relative proportions of resources to be represented where actual levels may not be known a priori. We evaluate our solution on a three-phase scenario of a real-time strategy game and compare the effectiveness of strategies that were purely human-devised, purely evolved, and those resulting from the human-evolution collaboration. The collaborative approach shows promising results in being able to find an optimum solution earlier.

Concepts from the Australian mathematics curriculum on fractions were used as core elements to design three computer games. In each game, the concepts were presented in the form of tangible puzzles, customized to a difficulty level based on student capability. The games were integrated into a single virtual game world, and a fantasy story was used to help build a compelling experience. Five Year 6 classes were used to evaluate the game over four weeks. Three of the classes were provided with the games, and two served as a control. Both the intervention and control groups also covered fractions in class as part of the regular teaching program, consisting of instructor led content combined with access to online resources and activities. Participants completed a diagnostic test before the trial, and again at the end, designed to assess competence in the fractions concepts targeted by the game. Results show that on average students who had access to the game in addition to the regular teaching scored higher than control group students. In particular, looking at just students who started with a lower level of fractions skills, greater improvement was seen in those that had access to the game.

The increasing elderly population presents a challenge on the resources of carers and assisted living communities. In this paper, we present an algorithm based around the Microsoft Kinect for monitoring activities of daily living. The system analyses the behaviour of occupants to provide carers with valuable observational data, and has the capacity to detect abnormal events in the home.