[Adapted from an investigation into the effect of ageing on expert memory with CHREST, UKCI 2007].

CHREST (Chunk Hierarchy and REtrieval STructures) is a cognitive architecture that models human perception, learning, memory, and problem solving. Influenced by the earlier EPAM model, it originated from modelling work on chess expertise.

See the Software page for implementations of CHREST.


The model combines low-level aspects of cognition (e.g., mechanisms monitoring information in short-term memory) with high-level aspects of cognition (e.g., use of strategies). It consists of perception facilities for interacting with the external world, short-term memory stores (in particular, visual and verbal memory stores), a long-term memory store, and associated mechanisms for problem solving. Short-term memory in CHREST contains references to chunks held in long-term memory, which are recognised through the discrimination network from information acquired by the perception system. (See the figure for an overview of the different parts of CHREST.)

Architecture of CHREST


Learning is seen as the acquisition of a network of nodes (chunks), which also become connected as a function of the similarity of their contents. Chunks can be seen as clusters of information that can be used as units of perception and meaning (the chunks in the simulations below will be fragments of chess positions). As in EPAM, long-term memory is represented as a discrimination network, which sorts and stores chunks.

Patterns that recur often in the environment make it possible for chunks to evolve into more complex data structures, known as templates. Templates are schema-like structures that have slots allowing values to be encoded rapidly.

Simulations are carried out by allowing the model to acquire knowledge by receiving stimuli representative of the domain under study. For example, during the learning phase of the chess simulations, the program incrementally acquires chunks and templates by scanning a large database of positions taken from master-level games. This makes it possible to create networks of various sizes, and so to simulate the behaviour of players of different skill levels. Taken together with the presence of time and capacity parameters, this enables CHREST to make unambiguous and quantitative predictions.


A significant aspect of CHREST is the importance it places on the perception process. Rather than passively collecting information from the environment, the process of information gathering is directed by knowledge already learned; this, in turn, affects the knowledge that is captured, resulting in complex emergent behaviour. In the case of chess experiments, perception is equated with eye movements (approximately corresponding to attention), which are directed by chunks held in memory and heuristics.


An important requirement of any model that is claimed to simulate human cognition is that it not assume any abilities exceeding those of a human. Thus, the parameters of the CHREST model are restricted to the human limits as understood by our current comprehension of human psychology. For example, by default the size of visual short-term memory is limited to four items, and the time required for moving the eye (known as a “saccade”), is set to 30 ms.

The majority of the variables in CHREST are time-related, so an internal clock is used to keep track of them. Each time an action is simulated by the system that is understood to take real time, such as mentally moving a piece, the clock is incremented by that period of time, as measured or estimated (note that this time representation is independent of the time taken to simulate the event; the actual processing time may be shorter or longer). Thus, time-restricted problems (such as the experiments we describe here) can be simulated.