Microsoft Word – 补充.docx
你 好 ， 我 已 经 写 完 了 reactive architecture, 现 在 只 要 求 完 成 deliberative
architecture 就好了。然后我把上课讲的一些 deliberative architecture 再补充一
下，有用的话就看一下。
Deliberative architectures
• in a deliberative architecture, percepts (or communication) give rise to goals—
representations of a state to be achieved
• the agent deliberates about how to achieve the goal – deliberation involves
(usually systematic) exploration of
alternative courses of action – a deliberative architecture typically includes
automatic generation
and comparison of alternatives
• result of deliberation is a representation of of the action(s) to be performed
The role of representations
• deliberation involves the manipulation of a model of the world and possible
courses of action, rather than the world itself
• requires the ability to represent actions and derive the consequences of actions
without actually performing them, e.g.:
– by remembering their effects in previous, similar situations – by reference to a
causal model of the world
Counterfactual representations
• to represent desired states and the consequences of actions:
– some states of the agent must be counterfactual in the sense of referring to
hypothetical future states (goals) or as yet unexecuted actions (plans)
– some of the basic operations of the architecture should generate such
counterfactual states
– such states must be influential in the choice of actions • to represent
hypothetical situations, a deliberative agent requires
representations with compositional semantics
我的理解就是说在这个 coursework 里，当这个 agent 做出选择的时候，有许多
路径：先去 station ,再去 refuel ,再去 well; 或者先去 refuel 再去 station 再去
well; 等等。 所以 deliberative architecture 就是能用一些算法 像 backward-
chaining algorithm 在之前给的 report 里写的，或者像 travelling-sale man
problem 里面的 TSP algorithm。然后就是能找到里面最好的路径，然后做出选
择。
之 前 提 供 的 report 里 面 用 的 是 hybrid architecture, 它 是 结 合 了 reactive
architecture 和 deliberative architecture，然后我把它的实现代码也附上了，这
个是之前的作业，和这次的稍有不同。我觉得它的实现太复杂了，我的只要
deliberative architecture 像上面那个图那样就行了。
下面是老师总结的一些设计时候的难点
Summary: design challenges
• what are the objectives of exploration?
• what are the constraints?
• what information can be/should be obtained?
• what is the “best” (simplest to program/most efficient/most
flexible/ …) strategy to obtain the information?
• is more than one strategy required – if so, how/when are these selected?
• how much exploration is “enough”?
还有在 explore 的时候：我也觉得下面那个（2.3）可能会好一点。