Using artificial intelligence to make decisions
Once you have collected and analysed your data, you have a transformed data set that represents a body of knowledge that you have extracted from your data. How can you make use of this knowledge?
You will need some form of decision-making process. These range from expert systems (the encoding of an experts knowledge), case-based reasoning (comparing with previous similar cases), decision trees (using predictable scenarios) through to neural networks, deep learning and other data mining methods.
There are many specific implementations of these and it is not necessary to dig deep into each. However, an overview of how they make, or help, a decision is useful:
Expert systems: these attempt to encode the knowledge of an expert and make similar decisions in similar cases. They come in two main varieties:
- Forward chaining: this is used to create a hypothesis from your data. For example: It is warm-blooded -> has feathers -> is black and white => it is a penguin.
- Backward chaining: this is used to confirm a hypothesis. For example: Is it a penguin? -> Is it warm-blooded? Yes -> does it have feathers? Yes -> is it black and white? Yes => It is a penguin.
- Decision trees: these can be generated automatically by Artificial Intelligence (AI) in many cases. We ask a question down each ‘branch’ of the ‘tree’ and separate out our data into ‘yes/no’ partitions until we have inseparable groups.
Neural networks and deep learning: these are ‘sub-symbolic’ techniques – in other words, the decisions they make cannot be represented using mathematical-based or language-based symbols. They will arrive at a decision but we cannot interpret the decision-making process. Because the reasoning for the decision cannot be understood, there are always questions about the reliability of decisions made in this way. The solutions arrived at by these AI systems can be drastic and are often controversial. This is illustrated in the Guardian article ‘Google’s solution to accidental algorithmic racism: ban gorillas’ (Hern 2018).
General learning algorithms:
- Supervised learning requires labelled data; the AI then learns to place similar data together in groups. Labelled data is not always available.
- Unsupervised learning attempts to discover natural groupings in data, without pre-labelled data. However, not all ‘groups’ are separate, especially if data is continuous. For example, how would you group ‘tall’ people and ‘short’ people?
- Static algorithms: these analyse a data set that doesn’t change – in other words, all the data to be analysed is available and none will be added later that will affect the decisions.
- Dynamic algorithms: these allow for analysis to be updated as further data arrives, without the need to re-run the complete analysis again. Groupings may update, move in the data space and merge.
- Evolving algorithms: these allow for analysis to be updated as further data arrives, similar to dynamic algorithms. However, evolving algorithms allow groups to not only move and merge, but also to divide, die out, or be created.
There are many algorithms in each of these categories. Some can be application-specific, others are more general but none are suitable to solve all of our AI requirements.
If you sign up for the rest of this Coventry University program, the next short course (Fundamental Machine Learning for AI) will illustrate how AI is used in these decision-making processes.
Think of an example of a decision you might make in everyday life and consider whether you make a ‘supervised’ or ‘unsupervised’ decision. This could be as simple as choosing a drink or deciding whether a purchase is too expensive.
Are there any circumstances in which the decision might be made in the other manner?
If you would like to learn more about the difference between symbolic AI and non symbolic AI you might want to read this article:
Bhatia, R. (2017) ‘Understanding the Difference between Symbolic AI & Non Symbolic AI’, Analytics India Magazine.
Hern, A. (2018) ‘Google’s Solution to Accidental Algorithmic Racism: Ban Gorillas’. Guardian [online] 12 January. available from https://www.theguardian.com/technology/2018/jan/12/google-racism-ban-gorilla-black-people [22 April 2020]
© Coventry University. CC BY-NC 4.0