Source Code for Module teamwork.policy.StochasticPolicy

 1  from teamwork.math.boltzmann import * 
 2   
 3  from LookaheadPolicy import * 
 4  from LookupAheadPolicy import * 
 5   
 6  import random 
 7 -class StochasticLookahead(LookaheadPolicy): 
 8      """A nondeterministic version of the lookahead-based policy, where 
 9      actions are selected with a probability that is a function of 
10      their expected values.""" 
11      # This policy's "temperature" constant for controlling the 
12      # probability distribution of actions.  Increasing values produce 
13      # a behavior closer to deterministic lookahead; decreasing values 
14      # produce a behavior closer to uniform randomness 
15      beta = 1. 
16       
17 -    def execute(self,state,choices=[],debug=Debugger(),depth=-1): 
18          """Returns a randomly selected action out of the available 
19          choices, with each action selected with a probability 
20          dependent on its relative expected value""" 
21          # Compute the EV of each option 
22          values = self.evaluateChoices(state=state,choices=choices, 
23                                        debug=debug,depth=depth) 
24          # Compute the probability distribution 
25          for option in values.values(): 
26              option['whole value'] = option['value'] 
27              option['value'] = option['value'].total().mean() 
28          self.computeDistribution(values) 
29          for option in values.values(): 
30              debug.message(7,'P(%s) = %4.3f' % (`option['decision']`, 
31                                                 option['probability'])) 
32          # Choose an action according to this distribution 
33          cutoff = random.random() 
34          total = 0. 
35          for option in values.values(): 
36              total += option['probability'] 
37              if total > cutoff: 
38                  break 
39          # Return the selected action 
40          action = option['decision'] 
41          del action['actor'] 
42          explanation = {'options':values, 
43                         'value':option['whole value'], 
44                         'decision':action, 
45                         'actor':state.name, 
46                         'breakdown':option['breakdown'], 
47                         'effect':option['effect'], 
48                         'differential':0.} 
49          debug.message(9,'%s selects %s with Prob %4.3f' \ 
50                        % (state.name,`action`,option['probability'])) 
51          return action,explanation 
52   
53 -    def computeDistribution(self,options): 
54          """Computes a probability distribution over the provided 
55          dictionary of action choices.  Each value in the dictionary 
56          must have a 'value' field containing a float.  This method 
57          computes a Boltzmann distribution based on these values and 
58          stores it in the 'probability' field of each entry.  Modify 
59          the 'beta' attribute on this object to vary the steepness of 
60          the distribution (0 is a uniform distribution, increasing 
61          values lead to deterministic behavior).  To use a different 
62          distribution altogether, simply override this method.""" 
63          return prob(options,self.beta*float(len(options))) 
64