home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!cs.utexas.edu!sun-barr!ames!agate!spool.mu.edu!yale.edu!jvnc.net!news.edu.tw!news!Net.nthu.edu.tw!News.nthu.edu.tw!dr788307
- From: dr788307@cs.nthu.edu.tw (dr78)
- Subject: Neural Reinforcement Learning?
- Message-ID: <1992Dec28.142832.16169%dr788307@cs.nthu.edu.tw>
- Sender: news@News.nthu.edu.tw (Net News)
- Organization: National Tsing Hua University (HsinChu)
- Date: Mon, 28 Dec 1992 14:28:32 GMT
- Lines: 29
-
- Sutton's Adaptive Heuristic Critic (AHC) and Watkin's Q-learning are
- two popular reinforcement learning mechanisms, and their original designs
- are based on storing utilities on a look-up table structure. In order to
- properly predict (interpolate) utilities of unvisited states and to
- compress the utility storage, many literature tried and reported progresses
- in neural implementation.
-
- However, this will introduce new problems: (1) updating the utility of a state
- may undesiredly cause a large change the utilities of other states. Besides,
- (2) the training set seems to be more or less conflict since we have many utilities
- needed to assign to a state in the updating process.
-
- If we just use the on-line backpropagation to train a MLP (feeding a pair
- of state and its new utility), the problem (1) and (2) will occur and (2)
- emphasizes the problem (1). If we use a batch backpropagation, the table storage
- will be necessary to keep a single utility of each state (to avoid the problem (2)).
- Obviously, this kind of neural implementation is guided by the symbolic knowledge
- and violates the original purpose of saving memory.
-
- To date, the reported results seem to be limited on small applications. Thus,
- the batch backpropagation seems to be the way which those researchers chose. Am
- I right? But, this seems to a bad neural implementation since the training set
- (stored in a table) is guided by symbolic knowledges.
-
- Any comments will be highly appreciated!!
-
- -------------------------------------------------------------------------------
- Hown-Wen Chen
- e-mail: dr788307@cs.nthu.edu.tw
-