Literature reviews for coursework rarely line up with what you are actually curious about. This one was assigned reading for the AI module of my degree: Chen, Jakeman, and Norton's 2008 survey paper, "Artificial Intelligence Techniques: An Introduction To Their Use For Modelling Environmental Systems," published in Mathematics and Computers in Simulation. The environmental modelling angle was never the draw for me. What I wanted out of it was a clean overview of the ten AI techniques it covers, the kind that could be lifted out of their original context and applied elsewhere. I also skipped past most of the mathematical formalism, since that is not where I am yet with this material; the goal here was breadth over depth, mapping each technique to something I already find interesting rather than absorbing the environmental case studies the paper was actually built around.
Case-Based Reasoning (CBR)
CBR solves a new problem by recalling similar problems solved before, on the assumption that similar problems have similar solutions. It retrieves the closest matching case from a case database, adapts it into a proposed solution, then revises that solution by simulation or by testing it against the real problem. Once a solution proves itself, it gets retained in the case base for future retrieval. The retrieve-adapt-revise-retain cycle maps closely onto retrieval-augmented generation: both pull the most relevant precedent from a store of prior examples before generating an answer for the current input.
Rule-Based Systems (RBS)
RBS come in two forms: forward chaining, which is data driven, and backward chaining, which is goal driven. Rules are structured as condition-action (if-then) pairs and fed into an inference engine. A pattern matcher decides which rules are relevant to the current state, and a rule applier decides which of those matched rules actually fire. Rules do not need to be precise either; several approaches exist for handling uncertainty within the rule set. This structure is a natural fit for expert systems, where domain knowledge gets encoded directly as rules rather than learned from data.
Artificial Neural Networks (ANNs)
ANNs model, loosely, how the brain processes information: many simple processing nodes standing in for neurons, interconnected through weighted links that play the role of synapses. Structurally it is a graph, with an input layer, one or more hidden layers, and typically a single output, though there can be many inputs and many hidden nodes feeding into it. The comparison that stuck with me is the black-box nature of it, not far from how LLMs get discussed today; you can inspect the architecture and the weights, but the path from input to output resists straightforward explanation. The paper splits ANN training into two learning types: supervised learning, which in this framing includes reinforcement learning (the network is told whether an output was correct without being handed the correct value directly), and unsupervised learning, where the network explores and categorises unlabelled input patterns on its own. Some systems combine both. ANNs are the technique of choice when a deterministic algorithm is not available for the problem at hand.
Genetic Algorithms (GAs)
GAs are a search technique modelled on natural selection: survival of the fittest, reproduction, and the rest of the evolutionary toolkit, applied algorithmically to a population of candidate solutions. The main strength is implicit parallelism, the ability to evaluate and evolve multiple candidate solutions at once instead of committing to one early. That property is useful when the "correct" solution is highly context dependent and no single deterministic answer really exists.
Cellular Automata (CA)
CA are dynamic models, discrete in time, space, and state. A grid of cells updates synchronously at each time step, with each cell's new state computed from its own current state and the states of its neighbours according to local rules. Despite the simplicity of the local rules, CA are useful for modelling complex physical phenomena such as fluid dynamics, where global patterns emerge from purely local interactions.
Fuzzy Systems (FS)
Fuzzy systems handle incomplete and imprecise data. Classical set theory is binary: an object either belongs to a set or it does not. Fuzzy sets replace that binary membership with a value between 0 and 1, which lets fuzzy models represent vague statements in a way that is closer to natural language than strict logic allows.
Multi-Agent Systems (MAS)
MAS are networks of agents that collaborate because no single agent can solve the problem alone, communicating through an Agent Communication Language (ACL). This is reminiscent of the Agent Communication Protocol used today for autonomous agent-to-agent communication. It sounds powerful in theory, but in practice, chaining agents together risks becoming a pipeline for feeding slop from one agent into the next; the overall system's effectiveness ends up bottlenecked by that compounding noise and by hallucinations from any single agent in the chain.
Swarm Intelligence (SI)
SI is essentially hivemind modelling. Ant Colony Optimisation is one of the two major algorithms here, and it brought to mind the well-known experiment where slime mold was used to model the Tokyo rail network; both are cases of efficient network structure emerging from simple, local decision rules. Particle Swarm Optimisation, the other major algorithm, takes its inspiration from schools of fish and flocks of birds instead.
Reinforcement Learning (RL)
RL is about an agent learning through experimentation, trial and error, rather than from labelled examples. The paper breaks an RL problem into three parts: the environment, the reinforcement function, and the value function. It is commonly used in robotics, and more broadly in systems where the goal is improving efficiency and outcomes over repeated interaction rather than fitting a fixed dataset.
Hybrid Systems
Hybrid systems combine multiple paradigms to gain their combined strengths and offset their individual weaknesses. The paper identifies three main types: sequential, auxiliary, and embedded. The result is not guaranteed to be positive; some combinations end up inheriting the weaknesses of their parent techniques while losing most of the strengths, which is a useful reminder that combining tools is not automatically additive.
What Carries Over
Reading with an eye for transferable ideas rather than environmental applications, a few sections mapped cleanly onto things I already spend time on. CBR's retrieve-adapt-revise-retain loop is a good description of what RAG does today. ANNs as opaque graphs of weighted connections is basically the same complaint people make about LLMs. The MAS section, for all its promise, is a reminder that multi-agent LLM pipelines carry the same risk on paper as in practice: chaining agents together does not automatically produce a better answer than one agent would, especially once errors start compounding across hops. None of this is what the paper set out to argue; it translates well past the specific environmental use cases it was written for.
Reflection on the Reading Process
This was assigned reading, and a survey paper written for a specific application domain is not naturally gripping cover to cover, especially while skipping past most of the mathematical detail I have not built up yet. What made it worth the time was treating it as a checklist rather than a narrative: ten named techniques, each with a clear enough description to know what it is, when it might apply, and what to search for if I need the details later. That is a more useful outcome for a literature review than trying to absorb environmental modelling specifics I am unlikely to ever use directly.
Conclusion
None of these ten techniques are new or cutting edge, and the paper does not claim otherwise; it is a survey, not original research. That is exactly what makes it useful as a reference. Case-based reasoning, rule-based systems, neural networks, genetic algorithms, cellular automata, fuzzy systems, multi-agent systems, swarm intelligence, reinforcement learning, and hybrid systems are all AI paradigms worth being able to name and place, independent of the environmental systems the paper was written to model. The literature review requirement is done, but the more durable outcome is a short list of techniques I can now recognise and reach for regardless of where I first read about them.