Journal Article

Explaining Evolutionary Agent-Based Models via Principled Simplification

Posted on March 16, 2022 by Journal Article Publication

Chloe M. Barnes, Abida Ghouri, Peter R. Lewis: Explaining Evolutionary Agent-Based Models via Principled Simplification. In: Artificial Life, vol. 27, no. 3, 2021.

Abstract

Understanding how evolutionary agents behave in complex environments is a challenging problem. Agents can be faced with complex fitness landscapes derived from multi-stage tasks, interaction with others, and limited environmental feedback. Agents that evolve to overcome these can sometimes access greater fitness, as a result of factors such as cooperation and tool use. However, it is often difficult to explain why evolutionary agents behave in certain ways, and what specific elements of the environment or task may influence the ability of evolution to find goal-achieving behaviours; even seemingly simple environments or tasks may contain features that affect agent evolution in unexpected ways. We explore principled simplification of evolutionary agent-based models, as a possible route to aiding their explainability. Using the River Crossing Task (RCT) as a case study, we draw on analysis in the Minimal River Crossing (RC-) Task testbed, which was designed to simplify the original task while keeping its key features. Using this method, we present new analysis concerning when agents evolve to successfully complete the RCT. We demonstrate that the RC- environment can be used to understand the effect that a cost to movement has on agent evolution, and that these findings can be generalised back to the original RCT. Then, we present new insight into the use of principled simplification in understanding evolutionary agents. We find evidence that behaviour dependent on features that survive simplification, such as problem structure, are amenable to prediction; while predicting behaviour dependent on features that are typically reduced in simplification, such as scale, can be invalid.


BibTeX (Download)

@article{Barnes2021ExplainingSimplification,
title = {Explaining Evolutionary Agent-Based Models via Principled Simplification},
author = {Chloe M. Barnes and Abida Ghouri and Peter R. Lewis},
doi = {10.1162/artl_a_00339},
year  = {2021},
date = {2021-03-16},
urldate = {2021-01-01},
journal = {Artificial Life},
volume = {27},
number = {3},
publisher = {MIT Press},
abstract = {Understanding how evolutionary agents behave in complex environments is a challenging problem. Agents can be faced with complex fitness landscapes derived from multi-stage tasks, interaction with others, and limited environmental feedback. Agents that evolve to overcome these can sometimes access greater fitness, as a result of factors such as cooperation and tool use. However, it is often difficult to explain why evolutionary agents behave in certain ways, and what specific elements of the environment or task may influence the ability of evolution to find goal-achieving behaviours; even seemingly simple environments or tasks may contain features that affect agent evolution in unexpected ways. We explore principled simplification of evolutionary agent-based models, as a possible route to aiding their explainability. Using the River Crossing Task (RCT) as a case study, we draw on analysis in the Minimal River Crossing (RC-) Task testbed, which was designed to simplify the original task while keeping its key features. Using this method, we present new analysis concerning when agents evolve to successfully complete the RCT. We demonstrate that the RC- environment can be used to understand the effect that a cost to movement has on agent evolution, and that these findings can be generalised back to the original RCT. Then, we present new insight into the use of principled simplification in understanding evolutionary agents. We find evidence that behaviour dependent on features that survive simplification, such as problem structure, are amenable to prediction; while predicting behaviour dependent on features that are typically reduced in simplification, such as scale, can be invalid.},
keywords = {Agent-Based Systems, ANNs, Artificial Life, Evolutionary Algorithms, Explainability, Neuroevolution, Principled Simplification, River Crossing},
pubstate = {published},
tppubtype = {article}
}

Fruit Quality and Defect Image Classification with Conditional GAN Data Augmentation

Posted on February 5, 2022 by Journal Article Publication

Jordan J. Bird, Chloe M. Barnes, Luis J. Manso, Anikó Ekárt, Diego R. Faria: Fruit Quality and Defect Image Classification with Conditional GAN Data Augmentation. In: Scientia Horticulturae, vol. 293, pp. 110684, 2022, ISSN: 0304-4238.

Abstract

Contemporary Artificial Intelligence technologies allow for the employment of Computer Vision to discern good crops from bad, providing a step in the pipeline of selecting healthy fruit from undesirable fruit, such as those which are mouldy or damaged. State-of-the-art works in the field report high accuracy results on small datasets (<1000 images), which are not representative of the population regarding real-world usage. The goals of this study are to further enable real-world usage by improving generalisation with data augmentation as well as to reduce overfitting and energy usage through model pruning. In this work, we suggest a machine learning pipeline that combines the ideas of fine-tuning, transfer learning, and generative model-based training data augmentation towards improving fruit quality image classification. A linear network topology search is performed to tune a VGG16 lemon quality classification model using a publicly-available dataset of 2690 images. We find that appending a 4096 neuron fully connected layer to the convolutional layers leads to an image classification accuracy of 83.77%. We then train a Conditional Generative Adversarial Network on the training data for 2000 epochs, and it learns to generate relatively realistic images. Grad-CAM analysis of the model trained on real photographs shows that the synthetic images can exhibit classifiable characteristics such as shape, mould, and gangrene. A higher image classification accuracy of 88.75% is then attained by augmenting the training with synthetic images, arguing that Conditional Generative Adversarial Networks have the ability to produce new data to alleviate issues of data scarcity. Finally, model pruning is performed via polynomial decay, where we find that the Conditional GAN-augmented classification network can retain 81.16% classification accuracy when compressed to 50% of its original size.


BibTeX (Download)

@article{Bird2022FruitAugmentation,
title = {Fruit Quality and Defect Image Classification with Conditional GAN Data Augmentation},
author = {Jordan J. Bird and Chloe M. Barnes and Luis J. Manso and Anikó Ekárt and Diego R. Faria},
doi = {10.1016/j.scienta.2021.110684},
issn = {0304-4238},
year  = {2022},
date = {2022-02-05},
urldate = {2022-01-01},
journal = {Scientia Horticulturae},
volume = {293},
pages = {110684},
abstract = {Contemporary Artificial Intelligence technologies allow for the employment of Computer Vision to discern good crops from bad, providing a step in the pipeline of selecting healthy fruit from undesirable fruit, such as those which are mouldy or damaged. State-of-the-art works in the field report high accuracy results on small datasets (<1000 images), which are not representative of the population regarding real-world usage. The goals of this study are to further enable real-world usage by improving generalisation with data augmentation as well as to reduce overfitting and energy usage through model pruning. In this work, we suggest a machine learning pipeline that combines the ideas of fine-tuning, transfer learning, and generative model-based training data augmentation towards improving fruit quality image classification. A linear network topology search is performed to tune a VGG16 lemon quality classification model using a publicly-available dataset of 2690 images. We find that appending a 4096 neuron fully connected layer to the convolutional layers leads to an image classification accuracy of 83.77%. We then train a Conditional Generative Adversarial Network on the training data for 2000 epochs, and it learns to generate relatively realistic images. Grad-CAM analysis of the model trained on real photographs shows that the synthetic images can exhibit classifiable characteristics such as shape, mould, and gangrene. A higher image classification accuracy of 88.75% is then attained by augmenting the training with synthetic images, arguing that Conditional Generative Adversarial Networks have the ability to produce new data to alleviate issues of data scarcity. Finally, model pruning is performed via polynomial decay, where we find that the Conditional GAN-augmented classification network can retain 81.16% classification accuracy when compressed to 50% of its original size.},
keywords = {CNNs, Data Augmentation, Fruit Quality, GANs, Image Classification},
pubstate = {published},
tppubtype = {article}
}

Behavioural Plasticity Can Help Evolving Agents in Dynamic Environments But at the Cost of Volatility

Posted on December 20, 2021 by Journal Article Publication

Chloe M. Barnes, Anikó Ekárt, Kai Olav Ellefsen, Kyrre Glette, Peter R. Lewis, Jim Tørresen: Behavioural Plasticity Can Help Evolving Agents in Dynamic Environments But at the Cost of Volatility. In: ACM Transactions on Autonomous Adaptive Systems, vol. 15, no. 4, 2021.

Abstract

Neural networks have been widely used in agent learning architectures; however, learnings for one task might nullify learnings for another. Behavioural plasticity enables humans and animals alike to respond to environmental changes without degrading learned knowledge; this can be achieved by regulating behaviour with neuromodulation—a biological process found in the brain. We demonstrate that by modulating activity-propagating signals, neurally trained agents evolving to solve tasks in dynamic environments that are prone to change can expect a significantly higher fitness than non-modulatory agents and also achieve their goals more often. Further, we show that while behavioural plasticity can help agents to achieve goals in these variable environments, this ability to overcome environmental changes with greater success comes at the cost of highly volatile evolution.


BibTeX (Download)

@article{Barnes2021BehaviouralVolatility,
title = {Behavioural Plasticity Can Help Evolving Agents in Dynamic Environments But at the Cost of Volatility},
author = {Chloe M. Barnes and Anikó Ekárt and Kai Olav Ellefsen and Kyrre Glette and Peter R. Lewis and Jim Tørresen},
doi = {10.1145/3487918},
year  = {2021},
date = {2021-12-20},
urldate = {2021-01-01},
journal = {ACM Transactions on Autonomous Adaptive Systems},
volume = {15},
number = {4},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
abstract = {Neural networks have been widely used in agent learning architectures; however, learnings for one task might nullify learnings for another. Behavioural plasticity enables humans and animals alike to respond to environmental changes without degrading learned knowledge; this can be achieved by regulating behaviour with neuromodulation—a biological process found in the brain. We demonstrate that by modulating activity-propagating signals, neurally trained agents evolving to solve tasks in dynamic environments that are prone to change can expect a significantly higher fitness than non-modulatory agents and also achieve their goals more often. Further, we show that while behavioural plasticity can help agents to achieve goals in these variable environments, this ability to overcome environmental changes with greater success comes at the cost of highly volatile evolution.},
keywords = {Agent-Based Systems, ANNs, Artificial Life, Behavioural Plasticity, Evolutionary Algorithms, Evolutionary Volatility, Neuroevolution, Neuromodulation, River Crossing},
pubstate = {published},
tppubtype = {article}
}

Beyond Goal-Rationality: Traditional Action Can Reduce Volatility in Socially Situated Agents

Posted on December 1, 2020 by Journal Article Publication

Chloe M. Barnes, Anikó Ekárt, Peter R. Lewis: Beyond Goal-Rationality: Traditional Action Can Reduce Volatility in Socially Situated Agents. In: Future Generation Computer Systems, vol. 113, pp. 579–596, 2020.

Abstract

Systems that pursue their own goals in shared environments can indirectly affect one another in unanticipated ways, such that the actions of other systems can interfere with goal-achievement. As humans have evolved to achieve goals despite interference from others in society, we thus endow socially situated agents with the capacity for social action as a means of mitigating interference in co-existing systems. We demonstrate that behavioural and evolutionary volatility caused by indirect interactions of goal-rational agents can be reduced by designing agents in a more socially-sensitive manner. We therefore challenge the assumption that designers of intelligent systems typically make, that goal-rationality is sufficient for achieving goals in shared environments.


BibTeX (Download)

@article{Barnes2020BeyondAgents,
title = {Beyond Goal-Rationality: Traditional Action Can Reduce Volatility in Socially Situated Agents},
author = {Chloe M. Barnes and Anikó Ekárt and Peter R. Lewis},
url = {https://www.sciencedirect.com/science/article/pii/S0167739X20304714},
doi = {10.1016/j.future.2020.07.033},
year  = {2020},
date = {2020-12-01},
urldate = {2020-01-01},
journal = {Future Generation Computer Systems},
volume = {113},
pages = {579--596},
abstract = {Systems that pursue their own goals in shared environments can indirectly affect one another in unanticipated ways, such that the actions of other systems can interfere with goal-achievement. As humans have evolved to achieve goals despite interference from others in society, we thus endow socially situated agents with the capacity for social action as a means of mitigating interference in co-existing systems. We demonstrate that behavioural and evolutionary volatility caused by indirect interactions of goal-rational agents can be reduced by designing agents in a more socially-sensitive manner. We therefore challenge the assumption that designers of intelligent systems typically make, that goal-rationality is sufficient for achieving goals in shared environments.},
keywords = {Agent-Based Systems, ANNs, Artificial Life, Evolutionary Volatility, Neuroevolution, River Crossing, Social Action},
pubstate = {published},
tppubtype = {article}
}

Country-level Pandemic Risk and Preparedness Classification based on COVID-19 Data: A Machine Learning Approach

Posted on October 10, 2020 by Journal Article Publication

Jordan J. Bird, Chloe M. Barnes, Cristiano Premebida, Anikó Ekárt, Diego R. Faria: Country-level Pandemic Risk and Preparedness Classification based on COVID-19 Data: A Machine Learning Approach. In: PLOS ONE, vol. 15, no. 10, pp. 1–20, 2020.

Abstract

In this work we present a three-stage Machine Learning strategy to country-level risk classification based on countries that are reporting COVID-19 information. A K% binning discretisation (K = 25) is used to create four risk groups of countries based on the risk of transmission (coronavirus cases per million population), risk of mortality (coronavirus deaths per million population), and risk of inability to test (coronavirus tests per million population). The four risk groups produced by K% binning are labelled as ‘low’, ‘medium-low’, ‘medium-high’, and ‘high’. Coronavirus-related data are then removed and the attributes for prediction of the three types of risk are given as the geopolitical and demographic data describing each country. Thus, the calculation of class label is based on coronavirus data but the input attributes are country-level information regardless of coronavirus data. The three four-class classification problems are then explored and benchmarked through leave-one-country-out cross validation to find the strongest model, producing a Stack of Gradient Boosting and Decision Tree algorithms for risk of transmission, a Stack of Support Vector Machine and Extra Trees for risk of mortality, and a Gradient Boosting algorithm for the risk of inability to test. It is noted that high risk for inability to test is often coupled with low risks for transmission and mortality, therefore the risk of inability to test should be interpreted first, before consideration is given to the predicted transmission and mortality risks. Finally, the approach is applied to more recent risk levels to data from September 2020 and weaker results are noted due to the growth of international collaboration detracting useful knowledge from country-level attributes which suggests that similar machine learning approaches are more useful prior to situations later unfolding.


BibTeX (Download)

@article{Bird2020Country-levelApproach,
title = {Country-level Pandemic Risk and Preparedness Classification based on COVID-19 Data: A Machine Learning Approach},
author = {Jordan J. Bird and Chloe M. Barnes and Cristiano Premebida and Anikó Ekárt and Diego R. Faria},
doi = {10.1371/journal.pone.0241332},
year  = {2020},
date = {2020-10-10},
urldate = {2020-01-01},
journal = {PLOS ONE},
volume = {15},
number = {10},
pages = {1--20},
publisher = {Public Library of Science},
abstract = {In this work we present a three-stage Machine Learning strategy to country-level risk classification based on countries that are reporting COVID-19 information. A K% binning discretisation (K = 25) is used to create four risk groups of countries based on the risk of transmission (coronavirus cases per million population), risk of mortality (coronavirus deaths per million population), and risk of inability to test (coronavirus tests per million population). The four risk groups produced by K% binning are labelled as ‘low’, ‘medium-low’, ‘medium-high’, and ‘high’. Coronavirus-related data are then removed and the attributes for prediction of the three types of risk are given as the geopolitical and demographic data describing each country. Thus, the calculation of class label is based on coronavirus data but the input attributes are country-level information regardless of coronavirus data. The three four-class classification problems are then explored and benchmarked through leave-one-country-out cross validation to find the strongest model, producing a Stack of Gradient Boosting and Decision Tree algorithms for risk of transmission, a Stack of Support Vector Machine and Extra Trees for risk of mortality, and a Gradient Boosting algorithm for the risk of inability to test. It is noted that high risk for inability to test is often coupled with low risks for transmission and mortality, therefore the risk of inability to test should be interpreted first, before consideration is given to the predicted transmission and mortality risks. Finally, the approach is applied to more recent risk levels to data from September 2020 and weaker results are noted due to the growth of international collaboration detracting useful knowledge from country-level attributes which suggests that similar machine learning approaches are more useful prior to situations later unfolding.},
keywords = {AI and Health, COVID-19, Machine Learning, Pandemic Risk Analysis},
pubstate = {published},
tppubtype = {article}
}