David Joslin

Agent-based simulation for software project planning

Estimates of task duration and resource requirements in software engineering are notoriously inaccurate, and as a result effective project management often must be very dynamic. In response to new information or revised estimates, it may be necessary to reassign resources, cancel optional tasks, etc. Project management tools that make projections while treating decisions about tasks and resource assignments as static will not yield realistic results. In this paper we describe some preliminary attempts to adapt a simulation-based planning algorithm developed for planning experimental activities of Mars rovers to the problem of planning for software project management. Simulation techniques offer the potential for modeling the way agents behave in project development, and the way a manager might adapt the project plan based on the project status at future points, resulting in a tool that more accurately reflects the realities of software project management

Simulation-based planning for planetary rover experiments

Time and resource limitations mean that current Mars rovers (and any future planetary rovers) cannot hope to achieve every desirable scientific goal. We must therefore select and plan for a subset of the possible experiments, maximizing some utility metric. The use of simulation in planning is appealing because of its potential for representing complex, realistic details about the rover and its environment. We demonstrate a planning algorithm that performs high-level planning in a space of plan strategies, rather than actual plans. In the current implementation, candidate strategies are evaluated by a simple simulation, and a genetic algorithm is to search for effective strategies. Preliminary results are encouraging, particularly the potential for modeling uncertainty about the time required to complete actions, and the ability to develop strategies that can deal with this uncertainty gracefully

Opportunistic Fitness Evaluation in a Genetic Algorithm for Civil Engineering Design Optimization

The process of large structure design in civil engineering relies primarily on trial and error, guided by experience. We apply genetic algorithms to search for valid designs (satisfying all design constraints), minimizing total weight. The fitness evaluation has two components. Evaluating the validity of a candidate solution is very expensive, but the total weight can be evaluated independently and relatively cheaply. We demonstrate two techniques for using the inexpensive quality evaluation to decide whether or not the expensive validity evaluation is worth the investment of time it requires. We also use operators that reflect domain expert knowledge about design improvement techniques in order to improve convergence.

Real realization: Dennett's real patterns versus Putnam's ubiquitous automata

Both Putnam and Searle have argued that that every abstract automaton is realized by every physical system, a claim that leads to a reductio argument against Cognitivism or Strong AI: if it is possible for a computer to be conscious by virtue of realizing some abstract automaton, then by Putnam’s theorem every physical system also realizes that automaton, and so every physical system is conscious—a conclusion few supporters of Strong AI would be willing to accept. Dennett has suggested a criterion of reverse engineering for identifying “real patterns,” and I argue that this approach is also very effective at identifying “real realizations.” I focus on examples of real-world implementations of complex automata because previous attempts at answering Putnam’s challenge have been overly restrictive, ruling out some realizations that are in fact paradigmatic examples of practical automaton realization. I also argue that some previous approaches have at the same time been overly lenient in accepting counter-intuitive realizations of trivial automata. I argue that the reverse engineering approach avoids both of these flaws. Moreover, Dennett’s approach allows us to recognize that some realizations are better than others, and the line between real realizations and non-realizations is not sharp.

Leveraging domain-expert knowledge in a genetic algorithm for civil engineering design optimization

Chromosome evaluation for the civil engineering truss optimization problem is very expensive, thus limiting the number of iterations we can afford to perform. To try to improve convergence we used several mutation operators modeled after strategies used by domain experts. We ran experiments with various combinations of the operators and found that the domain-expert knowledge significantly improves the performance of the GA.