To reconstruct the concept of a LCS again we will following the following general 'roadmap':
- Prepare a statistical model for the wood1 as well as wood1-similar environments (cf. Wilson 1994 [347]).
- Define a first set of simple tasks to be solved (this represents environment ).
- Prepare a purely random agent
.
- Show the performance of agent
for these tasks.
- Construct an enhanced agent
with a complete reactive behavior for every possible perception.
- Show the performance of this
compared to
.
- Construct an enhanced agent from Wilson 1994 [347] with GA-like learning based on an external fitness function.
- Show the performance of this compared to
and
.
- Construct an enhanced agent
with a 2-level memory and a built-in fitness function. The GA-like learning uses the built-in fitness function as well as the memory, which can show actions and their results from the past organized in different paths.
- Show the performance of this
compared to all the other preceding agent types.
- Construct an enhanced agent
with a 4-level memory and a built-in fitness function. The GA-like learning uses the built-in fitness function as well as the memory, which can show actions and their results from the past organized in different paths on the levels 1-2. On level 3 the memory represents non-language objects
as well as language objects . On level 4 the memory represents sign relations mapping language objects onto non-language objects and vice versa
.
- Show the performance of this
compared to all the other preceding agent types.
Gerd Doeben-Henisch
2012-03-31