Goal Context | Body
5.5 Alternate reasoning methods
5.5.1 Classical agent reasoning
Classical artificial intelligence systems are based on the symbolic representation and manipulation of knowledge for their decision-making process. The control of classical agent-like systems, such as SHRDLU, STRIPS and NOAH, were based on deliberative plan generation, where the problem-solving follows the sense-plan-act process, the planning problem was described in terms of the state of the world, the desired goal state and a set of operators to effect changes to the world. The knowledge bases maintained by these systems were both the agent’s internal model of the environment and the application’s simulation of the environment. The systems assumed that the agent had a complete and up to date view of the environment and that no changes occurred in the
75
environment outside of the control of the agent. Little emphasis was placed on the execution of actions so manipulating items in the environment was simply a task of altering a symbolic statement in the knowledge base. In addition the information about the environment was represented as a set of highly abstract, symbolic statements about the environment. Although the systems produced positive results in their environment they suffered the limitation of being less successful when applied to real world environments.
5.5.1.1 Deliberative agents
One of the more sophisticated deliberative agents was the Homer project (Vere & Bickmore 1990), which was an attempt to construct a complete socially aware rational agent that was able to function in a simulated dynamic environment. The environment called Seaworld simulated the activity around a small harbour (see figure 5.4) containing a number of objects such as docks, islets, fish and passing boats. The agent, called Homer, operated as an Autonomous Underwater Vehicle (AUV) able to sense, make plans, perform actions, communicate in a subset of English and reflect upon its activities in the environment.
Figure 5.4. The seaworld environment – (Courtesy of Vere & Bickmore 1990) STEVE> What is in front of you?
HOMER> A log.
STEVE> Do you own the log? HOMER> No I don’t.
STEVE> The log belongs to you. HOMER> Oh.
STEVE> Cows eat grass. HOMER> I know.
STEVE> Do you own the log now? HOMER> Yes I do.
76
The goal of the developers was to integrate the then technology to develop an autonomous intelligent agent. Homer did address some of the deficiencies of deliberative systems. The knowledge base for the agent and the environment were separate, the agent had limited sensory abilities so was only “aware” of its immediate surroundings and changes could be made to the world outside of the agents knowledge. To allow Homer to function in the environment its reasoning capabilities were built around specialised modules such as a temporal plan generator, an action executor, different types of agent memory for different tasks and a reflective processor. Homer also contained natural language processing modules for communication with human users, including being set goals to achieve and commenting on its activities.
Although Homer is only capable of deliberative processing it is able to react to changes in the environment by re-planning, making changes to the formulated plans in the agent memory to cope with the new information. Homer can be regarded as an advancement on the SHRDLU simulation system, where there was no distinction between the agent’s knowledge base and the environment. It was developed with the engineering goal of investigating the state of AI technology by producing a complete agent artefact rather than any particular contribution to research. However more recent research by Liu and Schubert use a similar planner and reasoning engine called ME (for Motivated Explorer) to research linguistic competence in self motivated intelligent agents (Liu & Schubert 2010).
5.5.1.2 Reactive agents
Completely reactive systems are able to rapidly process real world information that is often presented as a stream of data with very little abstraction from the environment. They are said to have advantages such as simplicity, economy and robustness against failure (Wooldridge 2002). However there are a number of difficulties, for example, as decisions are based on local information they are inherently short term and there is no principled methodology for building such agents.
77
Nils Nilsson proposed the Teleo-Reactive as an architecture for creating goal oriented reactive programs. The Teleo-Reactive (T-R) architecture is a reactive agent control system that directs an agent toward a goal in a manner that continuously takes into account the agent's changing perceptions of its environment. T-R programs are structured as a network of decision-making elements, processing directs an agent toward a goal in a manner that continuously takes into account the agent's changing perceptions of a dynamic environment to select the agent’s action. The programs are written in and interpreted by a production- rule-like condition-action language, where conditions may specify some detectable situation from the environment condition and actions specify agent behaviours. Although rule-based reasoning is generally associated with production systems they may also be used for plan generation and execution. Rules allow agent behaviours to be executed from simple operators rather than a library of pre-coded plans typical of BDI agents. In addition to continuous feedback, T-R programs support parameter binding and recursion. In addition, T-R programs are said to be intuitive and easy to write and are written in a form that is compatible with automatic planning and learning methods (Nilsson 1994). T-R programs have been used in the control of simulated agents and actual mobile robots.
Another example of a completely reactive agent is the subsumption architecture devised by Rodney Brooks, (Brooks 1991) who wanted to explore producing intelligence without the need for elaborate knowledge representation or reasoning. The idea of subsumption is to produce intelligent behaviour from a network of interacting stimuli-response subsystem modules, each of which controls a logically single or simple behaviour. The network of modules are organised into a fixed hierarchy where modules in lower layers represent primitive behaviours such as avoiding obstacles, which are able to override or subsume the behaviour effects from other modules at higher layers that govern more general tasks such as path following. In effect a subsumption architecture forms a software circuit analogous to an electronic circuit, where the operation at any one time is determined by the state of the inputs. There are two
78
mechanisms that allow modules to override the effects of other modules: suppression where the input to a module is blocked, hence preventing it producing a behaviour and inhibition where the output from a module is blocked. The reasoning for module behaviours are implemented as stimulus-response processes typically using condition-action rules and although computationally very simple the subsumption powered machines are capable of producing behaviours that would be regarded as sophisticated if produced by symbolic AI systems.