Applied Cognitive Systems

"Cognitive systems are natural or artificial information processing systems, including those responsible for perception, learning, reasoning and decision-making and for communication and action" (DTI Foresight initiative). This definition facilitated the conclusion that current artificial systems/robots are poor cognitive systems. A need was identified to improve devices that we use every day, including assistive technologies and to generate medical benefits. Our research aims to create flexible, robust and adaptive applied cognitive systems. There is much overlap and mutual benefit in the themes of CIRG, with a strong link to embodied machine intelligence here. ACS interacts with an environment, including virtual domains, seeking performance improvement through analogies with human/animal behaviours. EMI interacts with a physical environment and seeks performance improvements through any appropriate method. A common thread is the utilisation of Cybernetic feedback where interaction with the environment improves performance. Cognitive Systems research has been a pillar of Cybernetics since the coalescence of the research area in the 50s.

Research sub-areas include:

Learning

Genetics-Based Machine Learning is a family of optimisation techniques inspired by evolution that improve on a population of initially random solutions by selecting the most promising solutions and repeatedly "breeding" new solutions from them. As the generations progress, the population moves towards the best solution to a problem.

Learning Classifier Systems (LCS) are a population-based evolutionary technique, but rather than the genome representing a vector of numbers, it instead codes for a set of rules. Using such biologically-inspired methodology, members of CIRG have applied novel variants of LCS to solve problems from steel mill quality control (see the figure below), to multiplexing to robotic vacuum cleaner path planning, showing improvements over other existing methods.

Abstraction

The need for abstraction arose from the data-mining of rules in the steel industry through application of the genetics-based machine learning technique of Learning Classifier Systems, which utilise a Q-learning type update for reinforcement learning. It was noted that many rules had similar patterns. For example, there were many rules of the type 'if side guide setting < width, then poor quality product' due to different product widths. This resulted in a rule-base that was unnecessarily hard to interpret and slow to learn. The initial development of the abstraction method was based on the known problem of Connect4 due to its vast search space, temporal nature and available patterns. The novel Abstraction algorithm developed successfully improved the domain performance as higher-order abstracted rules replaced generalised state-action rules in a complex multi-step problem. It is hoped that this algorithm will help to fulfil the intended use of the LCS technique as a test bed for artificial cognitive processes. The figure shows a graph of percentage base rules versus abstracted rules (solid line) as training progresses (circle line).

Emotions

The importance of ‘emotions’ in control mechanisms for autonomous agents has been demonstrated using real and virtual robotic platforms. A novel agent architecture was developed to provide a foundation for ‘emotion’-based control. Instead of mapping states to actions, the novel system developed maps states to an analogue of emotions and then to states. This provided a non-linear, temporal control strategy that was non-deterministic and thus advantageous in tested exploratory domains. An appropriate test platform was created allowing real and virtual agents to coexist and allowed production of a number of emotional rules. The emotion-based architecture is shown to provide a number of benefits over conventional approaches, which include simpler behavioural programming and improved performance on complex exploration tasks. The two figures below show the results of conventional and emotional robot path planning. 

Value system

Artificial cognitive systems have had success in single objective, single domains where the worth of each action may be evaluated/estimated. However, if the system needs in to choose between multiple goals or select an action when the worth estimate is poor, e.g. due to long chains between current state and eventual payoff, then a value system will be required. There is current research interest in the game of Othello as strategy learning benefits fro m its value system being updated at each given state. Thus learning becomes a two-stage process; 1. learn the values of moves at each state, 2. learn the optimum policy of moves through the states. 

Memory

It is proposed that a biologically non-implausible model of working memory be created, incorporated into a general cognitive architecture, and embodied into an artificial agent (simulated and embodied in a real mobile robot), such that its interaction with a complex environment may be tested. Biological cognitive agents (e.g. humans, rats and other mammals) are located in the real world, so must act within it, whilst being constrained by it.